Skip to main content

Full text of "Behind deep blue : building the computer that defeated the world chess champion"

See other formats


BEHIND 



BLUE 

BUILDING THE COMPUTER THAT DEFEATED 
THI WOULD CHESS CHAMPION 



It 



FENG-HSIUNG HSU 



Behind Deep Blue 



Building the Computer That Defeated 
the World Chess Champion 



Feng-hsiung Hsu 



Princeton University Press 
Princeton and Oxford 



Copyright © 2002 by Princeton University Press 

Published by Princeton University Press, 
41 William Street, Princeton, New Jersey 08540 

In the United Kingdom: Princeton University Press, 
3 Market Place, Woodstock, Oxfordshire OX20 1SY 

All Rights Reserved 

ISBN 0-691-09065-3 (cloth : alk. paper) 

Library of Congress Cataloging-in-Publication Data 
has been applied for 

This book has been composed in Stone Serif and Stone Sans 
by Stephen I Pargeter, Banbury, Oxfordshire, UK 

Printed on acid-free paper. @ 

www.pupress.princeton.edu 

Printed in the United States of America 

10 987654321 



Contents 



Preface 

Acknowledgments 
Chess Notation 

Prologue 
Chapter 1 Show Time! 

Carnegie Mellon 
Chapter 2 An Office of Troublemakers 
Chapter 3 Taking the Plunge 
Chapter 4 The Chess Machine That Wasn't 
Chapter 5 The Race for First Machine Grandmaster 
Chapter 6 "Knock, Knock. Who's There?" 

Intermezzo 
Chapter 7 First Date with Histoij 

IBM 
Chapter 8 We Need a New Name 
Chapter 9 Bringing Up the Baby 
Chapter 10 A Living Mount Everest 
Chapter 1 1 Retooling 
Chapter 1 2 The Holy Grail 

Epilogue 
Chapter 13 Life after Chess 

Appendixes 
Appendix A A Lad from Taiwan 
Appendix B Selected Game Scores 
Appenux C Further Reading 



vn 
xi 
xv 



9 
20 
46 
69 
90 

105 

125 
143 
162 
186 
204 

261 

277 
292 
297 



Preface 



This book recounts my view of the adventure to create Deep Blue, 
the first computer to defeat the World Chess Champion in a seri- 
ous match. I started the project in 1985. Twelve years later, the 
adventure ended with Deep Blue setting a major milestone in 
human history and forever altering our view of how we would live 
with the computer. 

Chess machines like Deep Blue have a long history. The Turk, 
a "chess machine" built by Hungarian engineer Baron Wolfgang 
von Kempelen, first appeared at the Austrian court in 1769. Its 
most famous victim was probably Napoleon Bonaparte. The Turk, 
however, was a fake — a hidden human chess player made the 
moves for it. 

In the 1830s, Charles Babbage dreamed of building a chess 
machine to attract funding for his Analytic Engine (the first 
mechanical programmable computer, but never completed). He 
built a tic-tac-toe machine instead. A real chess machine was too 
hard, especially if you had to build it with only mechanical parts. 

The first electronic programmable computer, EN1AC, was com- 
pleted in 1946. Shortly after, in 1949, Claude Shannon gave a 
lecture that laid the foundation for modern chess computers. At 
that time, many renowned computer scientists believed that the 
Computer Chess Problem — creating a chess computer that could 
beat the World Champion — would be solved within a few years, 
but their predictions turned out to be overly optimistic. As time 
went on, and with no solution in sight, solving the Computer 
Chess Problem became a holy grail for computer scientists. By the 
early 1980s, over thirty years after Shannon's seminal work, the 



VII 



viii Preface 

best chess computers were playing at US national master level, still 
a long way from challenging the World Champion. 

My involvement with computer chess was by accident. In the 
spring of 1985, Professor H T Kung, my faculty advisor at Carnegie 
Mellon, informed me that my survival in graduate school was in 
doubt. A couple of weeks earlier, as a favor, I looked into a techni- 
cal problem for a leading computer chess expert in our 
department. He, the expert, did not like the solution that I came 
up with. Neither did I. After some thinking, I reached the conclu- 
sion that the real issue was with his approach. A radical new 
design, in part based on his competitors' ideas, was far superior. 
After he rejected my new solution, I told him bluntly that his 
approach was inferior and I had lost interest. Professor Kung 
explained that I should have been more tactful. He then urged me 
to write up a technical report and give a presentation to back up 
my assertion. I already had enough trouble from an earlier indis- 
cretion that almost got me kicked out of the school ... 

Thus began the strange and unexpected journey to create the 
most powerful chess-playing entity in the world, Deep Blue. 

Initially, I planned to retrace this journey along the same 
route as that taken in a favorite book of mine, The Double Helix: A 
Personal Account of the Discovery of the Structure of DNA by Nobel 
Laureate Dr James D Watson. As I was writing, however, I realized 
that the two books would be quite different in scope and nature. 
Dr Watson's book tells of one of the greatest scientific discoveries 
in the twentieth century. This book is more about an engineering 
quest than a scientific discovery. By nature, an engineering quest 
covers a richer fabric of life. The technical ideas behind the quest 
have first to be discovered and then carried to their logical com- 
pletion. The discovery may come from luck or inspiration, but the 
rest of the quest is a matter of perspiration and perseverance. The 
scope of the book, therefore, is closer to, say, Chuck Yeager's The 
Quest for Much One, his account of breaking the sound barrier. This 
book depicts the quest to exceed, even if only briefly, the chess- 
playing ability of the best human chess player on the planet. 

Comparatively, this book is perhaps more light-hearted than 
either of those mentioned above. One reason for this is the seem- 
ingly frivolous nature of the very idea of creating a computer that 
can defeat the World Chess Champion. Computer scientists are 
humans too, and they do like to have fun. The other reason is that 
I, and my writing, were greatly influenced by another favorite 



Preface ix 

book of mine, Surely You're Joking, Mr Feynman: Adventures of a 
Curious Character by Nobel Laureate Dr Richard P Feynman. From 
I)r Feynman's book, ( learned of his peculiar sense of humor, irre- 
pressible passion for life, and no-nonsense way of looking at 
things. But, more importantly, 1 learned of the value of being able 
to laugh at myself. Computer chess was a highly competitive 
research field — scientific rivalry was as strong and heated as in the 
I3NA research area. The final competition with Garry Kasparov, the 
World Chess Champion, was also an extremely serious affair. The 
ability to laugh at myself and not to take myself too seriously kept 
me sane during some very tense situations, arising first from the 
scientific rivalry and then from the two matches with Garry. If you 
can take a step back and see the larger picture, the most serious 
moments in life can also be the funniest. 

Many books cover the matches between World Chess 
Champion Garry Kasparov and Deep Blue, but most of them dwell 
on the "man versus machine" aspect. These books also tend to be 
instant chess books. The "man versus machine" angle apparently 
sells well for chess books, but it does not capture the true essence 
of the contest. The contest was really between men in two differ- 
ent roles: man as a performer and man as a toolmaker. Two 
matches were played between Garry and Deep Blue, with different 
results. In the 1996 match, man won as a performer, and in the 
1997 rematch, man won as a toolmaker. 

This book records the stories of the toolmakers, the people 
who built the machines. Chapter two, the early part of Chapter 
three and Appendix A are, however, about my life prior to the 
launching of the project. The stories in Chapter two and the early 
part of Chapter three have direct relevance to the project. The 
project existed only because events in these stories took place. 
Appendix A contains stories of my early life in Taiwan, and can be 
read as background material. I put this material in for two reasons. 
One is to answer the question of what kind of chess imbecile 
would decide to take on the world's greatest chess mind. The other 
reason is to provide you with clues about why things ended up the 
way they did. 

The stories in the book span two socially and culturally dif- 
ferent organizations, the Carnegie Mellon Computer Science 
Department and IBM Research. The first half of the book focuses 
on the project at Carnegie Mellon. During this period the basic 
technical ideas took shape, a new computer chess team was formed 



x Preface 

as a lark, and an intense scientific rivalry erupted. The rest of the 
book documents the evolution of the project at IBM: adjustment 
to the new environments, painful personnel changes, develop- 
ment of new chess machines and, of course, two intense and 
controversial matches between Garry and Deep Blue. The epilogue 
of the book covers post-match events, including Garry's apparent 
avoidance of a replay even though for two years after his defeat by 
Deep Blue, he had continuously thrown out a challenge for such a 
match. Garry lost the World Championship to Vladimir Kramnik 
in November 2000. Will this make a match against Kramnik a pos- 
sibility? It is not clear at this time of writing. 

A frequently-asked question about Deep Blue is, "Is it intelli- 
gent?" Garry's accusations of cheating both during and after the 
1997 match confirmed that Deep Blue passed the chess version of 
the Turing Test (a blind test to tell whether you are interacting with 
a human or a computer). But Deep Blue is not intelligent. It is only 
a finely-crafted tool that exhibits intelligent behavior in a limited 
domain. Garry lost the match, but he was the player with the real 
intelligence — Deep Blue would never be able to come up with the 
imaginative accusations. 

One final note: Deep Blue was a team effort. One of the dan- 
gers of writing a first-person account of a team effort is that the 
time and the energy put in by other members of the team can be 
under-represented. To reduce this problem, I have sought out team 
members' feedback on early drafts. If I still fail to credit someone's 
contribution properly, the fault is all mine. 



Acknowledgments 



The pioneering work of Joe Condon and Ken Thompson at Bell 
Labs, on the chess machine Belle, was the direct inspiration for the 
project. Ken also provided some of the endgame databases used by 
Deep Blue in both matches against Garry Kasparov. Lewis Stiller 
provided the other endgame databases. 

The Computer Science Department at Carnegie Mellon 
University and IBM Research were the two organizations that made 
the project possible. There are many other people from both 
organizations to whom I want to give special thanks: 

At Carnegie Mellon, 

• Deep Thought team members Thomas Anantharaman, 
Mike Browne, Murray Campbell, Peter Jansen, and Andreas 
Nowatzyk for their contributions and tolerance of my 
sometimes irrational behavior. 

• My advisor, Professor H T Kung, for his understanding and 
encouragement and for providing the funding to build 
Deep Thought. 

• Professor Raj Reddy and Professor Randy Bryant, for pro- 
viding additional financial assistance in the later phases of 
the Deep Thought project. 

• Lawrence Butcher and John Zsarnay, for providing much- 
needed technical support. 

• Various members of the Hitech team, but in particular Carl 
Ebeling and Gordon Goetsch, for valuable hints and inter- 
esting discussions. Dr Hans Berliner, who was more the 
royal opposition, provided the spur that kept us running 
forward. 

xi 



xii Acknowledgments 

• Other people who provided help and encouragement 
include Professor Roberto Bisiani, Kai-Fu Lee, Professor 
Tom Mitchell, Professor Danny Sleator, Professor Bob 
Sproull, and Professor Hide Tokuda. 

At IBM Research, 

• Deep Blue team members Murray Campbell and Joe 
Hoane, for their contribution and commitment to the 
project. In passing, I want to thank Murray's wife Gina and 
Joe's wife Elizabeth for their tolerance and understanding 
of the long work hours and the lost weekends. Several 
times, Murray and Joe had to bail me out of some techni- 
cal troubles associated with the bringing up of the Deep 
Blue chess chips, at the cost of their family lives. 

• The entire management chain, but especially our two 
managers, Randy Moulic and C J Tan, for their support and 
encouragement. 

• Jerry Brody, for providing the technical support for the 
team. 

Many outsiders have helped. Back in Deep Thought days, Larry 
Kaufman, Stuart Cracraft, Jim Gillogly, Frederic Friedel, Don 
Maddox and Dap Hartmann all provided help. Numerous IBMers, 
both in and outside IBM Research, were involved in hosting or pro- 
viding technical support for the two Deep Blue matches, as did 
people from ACM and TSI. I thank them all. 

Deep Blue would not be where it is without the help of many 
Grandmasters and other chess players who either played games 
against or trained with Deep Blue and its predecessors. Among the 
Grandmasters who had direct responsibilities at one time or anoth- 
er were Maxim Dlugy, Joel Benjamin, Miguel lllescas, Nick 
DeFirmian, and John Fedorowicz. Joel, of course, was also our 
chess advisor for both Deep Blue matches. Grandmaster Robert 
Byrne, the New York Times chess columnist, has been a good per- 
sonal friend who dropped in occasionally for a few games. Besides 
the Grandmasters who worked on the project, Robert was the one 
who provided the most input. Finally, without the participation of 
Garry Kasparov, the World Chess Champion, we would never have 
found out Deep Blue's real strength. 

Many people have reviewed earlier drafts of this book and 
made suggestions for improvements. Dennis Allison, Thomas 



Acknowledgments xiii 

Anantharaman, Mike Browne, Robert Byrne, Murray Campbell, 
Yih-Farn Robin Chen, Joe Hoane, Robert Hyatt, Howard Landman, 
Jim Loy, Timothy Mann, Andreas Nowatzyk, George Paul, 
Jonathan Schaeffer, Danny Sleator, C J Tan, Ken Thompson, and an 
anonymous reviewer all provided valuable feedback. My agent, 
William Clark, contributed an alternative view from his unique 
perspective. 

Some of the photos in the book are from my private collec- 
tion. Many of the match photos come from the IBM collection. 
Monty Newborn, Andreas Nowatzyk, and Danny Sleator provided 
the balance. 

Last but not least, I would like to thank David Ireland, my edi- 
tor, for refining and polishing my writing. In the end, however, 
any mistakes and omissions are solely my responsibility. 



Chess Notation 



This is not a chess book — there is no real chess analysis here. I have 
tried to avoid using chess notations or chess diagrams as much as 
possible. However, some of the controversies, especially those from 
the second Deep Blue match, centered on the chess moves played 
by Deep Blue. In theory, you could follow the controversies simply 
by treating the chess moves as labels, but it does no harm to know 
what the moves really mean. The chapters on the two Deep Blue 
matches contain chess diagrams to make it easier for you to visu- 
alize the game situations — you can treat them as though they are 
photos of the chessboard. Overleaf is a short description of the 
algebraic chess notation used in the book. 

The diagram shows the names for the eight rows (also known 
as ranks) and the eight columns (also known as files) of the chess- 
board. The rows are labeled 1 to 8 from the bottom row to the top. 
The columns are labeled a to h from the leftmost column to the 
rightmost. The name of each square of the chessboard is based on 
the column and the row that it is on. Therefore, the lower right 
corner square of the chessboard is known as hi, and the upper left 
corner square is known as a8. The chessboard itself can be divided 
vertically into two equal halves: the queenside (a, b, c, and d file) 
and the kingside (e, f, g, and h file). The names for the two halves 
come from the position of kings and queens on the starting chess- 
board. 

Not counting the pawn, there are five types of pieces: knight, 
bishop, rook, queen, and king. The pieces can be colored white or 
black. On the starting chessboard, white pawns are on the second 
rank and black pawns are on the seventh rank. The pieces on the 
first and the eighth rank are, from a file to h file, rook, knight, 

xv 



xvi Chess Notation 

bishop, queen, king, bishop, knight, and rook. The first rank pieces 
are white and the eighth rank pieces are black. The player with the 
white pieces is referred to as White, and the player with the black 
pieces, Black. 









w////a wm 'w& y& 



: y/M. W, HI 

I... w 

I 



1 



in 

''<"'" SffffrS* Sjr*, rf, /S, , < . ,, 



abcdefgh 

The starting chessboard, with labeled coordinates. 

Pawn moves are represented by the destination square of the 
pawn move. The move e4 is a pawn move to the e4 square, i'awn 
captures are represented by the name of the file that the pawn 
came from, followed by the destination square. An optional sym- 
bol 'x' can be placed between the file name and square name. 
Thus, the pawn capture axb4 uses an a-file pawn to capture a piece 
on the b4 square. 

Knight moves are represented by the piece symbol 'N' fol- 
lowed by the destination square. The move Nf3 is a move that 
places a knight on the f3 square. When there is more than one 
knight that can move to the same destination square, the name of 
the from-file or the name of the from-rank might be added before 
the name of the destination square. The move N2f3 would mean 
moving the knight on the second rank to the f3 square. If the 
knight move is a capture, then the optional capture symbol 'x' 
could be added before the destination square. The move NxgS 
therefore is a knight move that captures a piece on the g5 square. 

Bishop moves, rook moves, queen moves, and king moves fol- 



Chess Notation xvii 

low the same format as the knight moves but with the piece sym- 
bol replaced by 'B', 'R', 'Q' and 'K' respectively. The move Qa4 
therefore means moving the queen to the a4 square. 

The symbol '+' following a move means the move is a check; 
Rc4+ means rook to c4 and giving a check. 

Pawn promotions are represented by the pawn's move fol- 
lowed by the symbol for the piece to which the pawn is promoted. 
Therefore g8Q means pawn to g8 and promoted to a queen. 

Castling moves are represented by O-O and O-O-O, which 
mean castling kingside and castling queenside respectively. 

The move annotation symbols '!', '!!', '?', '??' mean "a good 
move", "an excellent move", "a questionable move" and "a dubi- 
ous move" respectively. So h5! means "pawn to h5, a good move". 



i. 



Prologue 



4) 



CHAPTER 1 



Show Time! 



In late April 1997, posters for an unusual chess event were appear- 
ing on the streets of New York. They showed a somber and 
pondering gentleman in his early 30s peering over a chess set at 
the viewers. The small caption under his chin said, "How do you 
make a computer blink?" The gentleman on the poster was the 
World Chess Champion Garry Kasparov, possibly the strongest 
chess player who has ever lived. 

Off the street, in the basement of the Equitable Building, I was 
staring at the blank screens in an empty auditorium. In a few days, 
the auditorium would be filled with an overflowing crowd; TV cam- 
eras would be entrenched at vantage locations and the three huge 
projection screens at the front would come to life. The left screen 
would be showing a live image from a TV studio on the 35th floor 
of the building, serving as the game room. The live image would 
usually show the two contestants sitting across a specially designed 
playing table. The contestant on the left would be Garry Kasparov. 
The contestant on the other side would be one of my two col- 
leagues, Murray Campbell and Joe Hoane, or me. Garry's real 
opponent was the chess computer, Deep Blue, that the three of us 
had designed and programmed. During the games we acted merely 
as extensions of Deep Blue and made moves for it on the physical 
chessboard. In the auditorium itself, three chess commentators, 
sometimes with a guest commentator or two, would be using the 
center screen to show their analysis of the ongoing game. The right 
screen would be displaying the overhead shot of the chessboard. 
This way, the audience in the auditorium would have a clear view 
of the present game position. 

It had taken me almost twelve years to reach this point. When 



4 Chapter 1 

I started, Garry was not the World Champion; it was a few months 
yet before he was crowned. For the past eleven years, since 1 986, 
my partners and I had been building successively more powerful 
chess computers. Our eventual goal was to beat the World Chess 
Champion, whoever he or she was. 

Before us, many pioneers, some famous and some not so 
famous, had made their contributions to the "Computer Chess 
Problem". In 1949, Claude Shannon made his proposal on how to 
program a computer to play chess. Since then, thousands of com- 
puter scientists, engineers, hobbyists, chess players, and even 
commercial organizations had worked on the problem. Some 
wanted to use chess as an experimental tool to find out how 
human intelligence worked. "If one could devise a successful chess 
machine, one would seem to have penetrated to the core of human 
intellectual endeavor," said Allen Newell, Cliff Shaw and Herbert 
Simon in one of the early computer chess papers. Other people 
viewed chess as a clear-cut, well-defined example of a complex 
problem. "Solving" chess could conceivably provide new tech- 
niques to solve other complex problems. The commercial entities 
did it for profit, of course, and some people, especially the hobby- 
ists, did it just for fun. 

We approached the problem from a different direction. We, or 
at least I, viewed the problem as a purely engineering one. Since 
the late 1970s, it had been established that chess computers 
became stronger as their hardware speed increased. By 1985, when 
1 started my small project that eventually become Deep Blue, the 
extrapolation from the experimental data indicated that a one 
thousandfold increase in hardware speed might be sufficient to 
produce a World Champion-class chess machine. Our project 
began with a simple goal, namely, to find out whether a massive 
increase in hardware speed would be sufficient to "solve" the 
Computer Chess Problem. Building this "Mother of all Chess 
Machines" was an interesting problem by itself. Of course, it would 
be an added bonus if our machine could indeed defeat the World 
Champion. 

The previous version of Deep Blue, lost a match to Garry 
Kasparov in Philadelphia in 1996. But two-thirds of the way into 
that match, we had played to a tie with Kasparov. That old version 
of Deep Blue was already faster than the machine that 1 conjec- 
tured in 1985, and yet it was not enough. There was more to 
solving the Computer Chess Problem than just increasing the 



Prologue 5 

hardware speed. Since that match, we rebuilt Deep Blue from 
scratch, going through every match problem we had and engaging 
Grandmasters extensively in our preparations. Somehow, all the 
work caused Grandmaster Joel Benjamin, our chess advisor, and 
one of the best chess players in the US, to say, "You know, some- 
times Deep Blue plays chess." Joel could no longer distinguish with 
certainty Deep Blue's moves from the moves played by the top 
Grandmasters. 

The press covered this new match with much anticipation. If 
the new Deep Blue won the match, then it would be a momentous 
occasion in the long history of men as toolmakers. It would also be 
the completion of a long-sought-after milestone for computer sci- 
entists and artificial intelligence researchers. It was almost certain 
that this match would be bigger than any World Chess 
Championship match, with possibly the sole exception of the 
Fischer vs. Spassky match in 1972. If we did win, perhaps not even 
that Fischer vs. Spassky match would compare. 

The new Deep Blue was much improved, but would it be 
enough? Would the journey begun by my partners and me so 
many years ago finally be over? 



Carnegie Mellon 



CHAPTER 2 



An Office of Troublemakers 



"Your office is an office of troublemakers!" Kung exclaimed, with a 
resigned look. Kung was Professor H T Kung, my faculty advisor at 
Carnegie Mellon University in Pittsburgh. He was a tall, well-built 
man in his 40s. I was a graduate student in the Computer Science 
(CS) Department. We had just had a long private conversation in 
his office. 

It was the autumn of 1987. Professor Kung had been my advi- 
sor during my entire stay at Carnegie Mellon. I had a great deal of 
respect for him, not just because he was a good advisor to me, but 
also because of the deeds that he did in private. Professor Kung was 
a direct descendent of Confucius and took great pride in doing 
whatever he could to assist the technical and economic develop- 
ments of Chinese societies around the world. I was not having 
trouble with Professor Kung. My troubles were with another facul- 
ty member in the department. 

At the time of the conversation, there were two computer 
chess teams at Carnegie Mellon. Dr Hans Berliner headed the 
Hitech team. The competing ChipTest team was a group of free- 
willed, mostly unsupervised graduate students, of which i was one. 
I had some light faculty supervision from Professor Kung, while 
my team members worked on ChipTest almost purely for fun. The 
competition between the two teams had been friendly, but it was 
about to turn into an intense rivalry. 

In hindsight, the rivalry was inevitable. There had been some 
friction between the two teams. The ChipTest team was not happy 
with the frequent and incorrect public portrayal that Dr Berliner 
was supervising us or that we were his students. To a great extent 
this portrayal was the result of Dr Berliner's existing fame in com- 



1 Chapter 2 

puter chess, which led reporters automatically to assume ' that he 
led all computer chess efforts at Carnegie Mellon. This characteri- 
zation was somewhat insulting to the ChipTest team, but it alone 
would not have caused the immediate escalation of the con- 
tention. 

The escalation was triggered by, of all things, an invitation for 
the ChipTest team to play in a chess tournament. The Hitech team 
happened to be going to the same tournament and Berliner had 
made his case to Kung on why ChipTest should skip the event 2 . 

My conversation with Professor Kung was precisely on what 
to do with the chess invitation. Kung was supportive of the 
ChipTest team and saw no reason why we should not play. 
However, to avoid further problems, the ChipTest team had 
already decided to honor Dr Berliner's suggestion prior to the con- 
versation. 

Professor Kung was relieved that the issue seemed to be 
resolved, but when he made his remark about the "troublemakers", 
he must have had a sense of deja vu. Two years before, in 1985, we 
had had another conversation, also on an issue related to Dr 
Berliner. That other conversation was a critical part of the chain of 
events that led to the creation of ChipTest and, eventually, Deep 
Blue. 

The real seed of the Deep Blue project can be traced back even 
further. It all began with Professor Kung's "troublemakers" ... 

First Year Graduate Students 

The Computer Science (CS) Department at Carnegie Mellon was 
the Department of the university. Several others ranked among the 
top ten in their disciplines, but CS was ranked among the top 
three, the other two being the ones at MIT and Stanford. 
Depending on personal preferences, some would even say that the 
CS Department at Carnegie Mellon was the best. The Department 
had its own unique atmosphere, and the graduate students had a 
great deal more say in their own research directions than at other 
schools. The fact that the Department offered only a PhD program 

1 This incorrect assumption persists to this day. Almost all the chess books on the 
Deep Blue matches state wrongly that Berliner was the faculty advisor of the 
ChipTest/Deep Thought team. 

2 The section An Invitation from California in Chapter 4 details how this whole inci- 
dent played out. 



An Office of Troublemakers 1 1 

(no BS or MS program) probably gave the faculty the chance to 
experiment and come up with the unique system. 

In part because of the unusual system and in part because of 
the diverse background of the incoming students, the entire first 
month of each fall semester was designated the "Immigration 
Course". For the incoming students, it was a time to learn about 
the environments, tools, projects, and people, and generally to 
have a good time. The older students would spend time demon- 
strating their projects, helping the new students to adapt, and of 
course, generally having a good time themselves. Near the end of 
the Immigration Course, the annual Department "Reception" 
would be held, usually at some fancy location in Pittsburgh. I have 
a fond memory of pigging out at the Reception as a starving grad- 
uate student. The Immigration Course was also the time the 
"Marriage" process took place. During the Immigration Course, the 
new students sought out the faculty whose research interested 
them, and decided whom they wanted as their faculty advisors. 
The faculty, of course, also observed the new students, deciding 
whether they wanted a particular student. As in real life, the 
Marriages were not final, and "divorces" between faculty members 
and students did occur sometimes, made possible partly by the fact 
that the funding sources for the students were not necessarily tied 
to specific projects. Faculty members controlled their advisees 
mainly through the respect that their advisees had for them. Other 
than course work, faculty members had very little hold over non- 
advisee students. 

In 1982, the year I "immigrated" into the Department, the 
acceptance rate of the admitted students was higher than expect- 
ed, and there was a space crunch. Normally when offices were 
assigned, it was considered preferable to mingle the incoming stu- 
dents with older students. As a result of the space crunch I ended 
up in an empty office together with three other incoming students, 
Mike Browne, Andreas Nowatzyk and Tony Stentz. In the end, the 
four of us came out all right, but along the way the faculty proba- 
bly wished that they had done otherwise. The next year, Tony 
moved to another office, and for several years only Mike, Andreas 
and I occupied the office, becoming close friends. Professor Kung's 
remark on troublemakers probably did not include Tony. 

Mike was probably the first American with whom I had a real 
conversation, which took place on my first day on the Carnegie 
Mellon campus. It was also only my third day in the United States. 



12 Chapter 2 

When I opened the door to my newly-assigned office, I found a 
bearded man inside. This was Mike. Because of his beard, I could 
not tell quite how old he was, but he turned out to be one of the 
younger students in the incoming class. Born in New Jersey, but 
growing up in Allentown, Pennsylvania, Mike finished his under- 
graduate degree at Lehigh University in just three years. In our 
conversation Mike described himself as a "hacker", which he 
defined as someone who could churn out computer programs. 
That is, someone who creates, although in a specialized domain. 
This definition is still in use today, although the more common 
definition used by the media is someone who damages or exploits 
computer systems through virtual means. The newer definition 
describes a destructive person. Mike was obviously very smart, but 
he was also surprisingly self-effacing. He attributed his accelerated 
college schooling entirely to his ability to do well in multiple- 
choice tests. "I just happened to be good at taking multiple-choice 
tests," he said, suggesting that he had no other useful talent. 

While Mike was one of the younger incoming students, 
Andreas and I were among the older ones. In my case, this was 
because in Taiwan, where I came from, boys were required to serve 
two years in the military after college 3 . Andreas was older because 
he had spent one extra year to get two tough undergraduate 
degrees: physics and computer science. He hailed from Germany, 
was a graduate of the University of Hamburg, and had been one 
year into his PhD program in Germany before he decided to make 
the move to Carnegie Mellon. Mike was an extremely good pro- 
grammer, and Andreas was not too shabby either. Besides being a 
top-notch programmer, Andreas turned out to be a man with 
many technical skills. He never ceased to amaze me while we were 
at Carnegie Mellon. I had a dogged belief that I could be good at 
anything technical if I put my mind to it. Andreas seemed to be 
already good at everything technical. 

Both Mike and I were somewhat playful in our early years at 
Carnegie Mellon, while Andreas was all business from day one. 
Andreas and I did not really become close friends until much later 
when we found that we shared many common technical interests. 
The friendship, however, was not really symmetrical. I can remem- 
ber many instances of Andreas helping me in areas where I was 

J See Sidestepping Authorities in A Lad in Taiwan, Appendix A, for a discussion on 
the military service in Taiwan. 



An Office of Troublemakers 1 3 

lacking. But he was so self-sufficient that I don't recall ever doing 
anything to in any way help him. 

Andreas, Mike, and I all had some prior contact with chess, or 
computer chess, before arriving in Carnegie Mellon. Mike used to 
be a serious tournament chess player when he was in high school. 
Andreas did not play chess seriously, but he had been keenly inter- 
ested in computer chess. He had once helped to referee a World 
Microcomputer Chess Championship tournament in Germany 
while an undergraduate student. My earlier contacts with comput- 
er chess were mainly from reading technical journals and books. 
During our immigration both Andreas and I checked out the exist- 
ing computer chess research in the Department. We were both 
interested in Very Large Scale Integration (VLSI) chip design and 
Carl Ebeling, a graduate student in the VLSI group, was designing 
a VLSI-based chess machine. Neither Andreas nor I were sufficient- 
ly interested to do anything with computer chess. At the time I 
thought computer chess still had a long way to go and I did not see 
how it would end. Moreover, I was really interested in doing some- 
thing practical. 

Professor Kung was the leading faculty member in the VLSI 
group and the obvious choice for both Andreas and me. Mike 
decided to work with a professor in Artificial Intelligence (AI). 

Andreas had a very low opinion of AI research in general. 
Mike did not share the same opinion at first but, after he split with 
his first advisor, both he and Andreas referred to AI as bullshit. I 
did not go to the same extreme, but I had seen some so-called 
research in AI that really deserved the bullshit label. The largest 
group by far in the Department was the AI group, including almost 
half of the students. The majority view toward AI in our office was 
clearly the minority view in the Department. 

Despite our mutual low opinion of AI, over the years Mike, 
Andreas and I had several conversations about computer chess. 
One of the conversations was about Belle, the top chess machine 
from the early 1980s. While Andreas was still at the University of 
Hamburg he had attended a talk by Ken Thompson, one of the 
Belle designers. In the talk, Ken described experiments that showed 
Belle playing more strongly when given more time. I posed the 
question of what would happen if a Belle-like machine were, say, 
one thousand times faster. None of us believed that such a 
machine would be sufficient to beat the World Champion, but it 
might be close. Anyway, when we had the conversation, none of 



14 Chapter 2 

us realized that I would drag both Andreas and Mike into a com- 
puter chess project trying to answer that very question. 

Professor Kung's remark about troublemakers perhaps includ- 
ed Andreas as, after a couple of years, Andreas and he had a 
falling-out. I don't know how bad Mike's split was, but it might 
have been bad enough for Professor Kung to have noticed at the 
"Black Friday" faculty meeting held at the end of each semester. 
Andreas could well have been high on Professor Kung's list of 
"troublemakers", but I probably took the top slot since I was the 
source of a lot of other troubles. Some of the troubles I caused 
resulted from immature pranks. Other troubles were not entirely 
my doing, as you will see in the next few chapters. 

The pranks that I pulled in the early years had two profound 
effects. The first was that I gained a reputation of a questionable 
nature among the students. This reputation ironically was helpful 
in recruiting other students to work on the chess project. (Of 
course, recruiting others was just the first step. In the end, the proj- 
ect had to be interesting in its own right and I had to treat my 
friends properly.) 

The second effect was to place me on the endangered species 
list at the Department. My first prank, mentioned in the next sec- 
tion, did not win me any friends at the Black Friday faculty 
meeting. My second prank almost got me kicked out of the school 
and could have landed me in jail if it had taken place a few years 
later. 

A License to Make Mischief 

Taiwanese students were supposed to pay strict attention to school 
rules and, while I was still in Taiwan, I was always fascinated by 
stories of students pulling ingenious April Fool's pranks on US col- 
lege campuses. 

During the Immigration Course, one of the sessions attended 
by everyone was the Department Introduction where we all 
squeezed into a large classroom and introduced ourselves to the 
rest of the Department. After introducing herself to the incoming 
students, Sharon Burks, who was the Department Administration 
Assistant at the time 4, jokingly advised whoever had forged an 

4 Sharon is now the Associate Department Head of the Computer Science 
Department and the Assistant Dean for Academic Affairs of the School of 
Computer Science at Carnegie Mellon. 



An Office of Troublemakers 15 

electronic post from her on April 1 to watch out. Surprised to hear 
of such insubordination from the older students, I said to myself, 
"This is neat. So we have a license to make mischief." 

My first year at Carnegie Mellon was a busy one. 1 wanted to 
pass all the core qualifiers as fast as ! could. The two years spent in 
military service in Taiwan had made me rusty so I also had a lot of 
ground to catch up. After the first semester, I was in decent shape 
school-wise, and my first April Fool's day at Carnegie Mellon was 
coming up. 

Back in the early 1980s, even in a top CS graduate school like 
Carnegie Mellon, the graduate students were only guaranteed 
access to text-only terminals in their offices. Some of us had office 
workstations that could display black-and-white graphics, and 
there were a few other workstations for public use in the Terminal 
Room. But most of our work was still done on terminals connect- 
ed to then state-of-the-art minicomputers. The minicomputers 
were less powerful than the personal microcomputers that we can 
buy today for less than $300 (which, by the way, was not enough 
to buy even one of the terminals that we used). Most of the mini- 
computers ran the UNIX operating system 5 . UNIX is widely used 
today in areas where computer crashes caused by the operating sys- 
tem are not tolerable. 

UNIX allows you to "talk" to another user on the same com- 
puter, assuming that the other user grants the privilege. Most users 
grant this privilege by default. When you "talk" to the other user, 
UNIX writes your typed characters directly onto the other user's 
terminal. Normally this would be harmless, other than possibly 
annoying the hell out of the other user if you get obnoxious. There 
is, however, a way to wreak havoc when the computer terminal 
used by the other user is somewhat "intelligent". The smarter 
computer terminals usually will accept command strings that the 
computer can use to move the cursor (the indicator on the screen 
which highlights where the new character is to appear) and to read 
back the (x, y) location of the cursor. By sending appropriate com- 
mand strings, you can move the cursor to a particular screen 
location and then cause the terminal to send out the location of 
the cursor back to your victim's computer. Since the computer is 

5 The operating system can be viewed as the basic program controlling all hard- 
ware aspects of a computer. Normal programs indirectly access the system 
hardware through the operating system. 



16 Chapter 2 

very likely expecting a user command, it usually appears as if the 
victim has entered the cursor location as a command. Presto, you 
have caused your victim to execute a command of your own 
choosing. This was the basic idea behind my April Fool's prank in 
1983. 

On March 31,1 had a conversation with a colleague who then 
got involved. Our plan was the following: I was to write two pro- 
grams. The first program would randomly select victims and send 
out appropriate strings to their terminals. The strings would then 
cause the victim's terminal to send out a command that executed 
my second program. The second program would post a random, 
possibly inappropriate, message on the Department electronic bul- 
letin board from the victim's account. My colleague was to come 
up with the collection of random messages for the posts, and he 
did a great job. The hardest technical part of the whole thing was 
to figure out a way that the victims would not know what had 
happened, even when they were in front of their terminals. The 
cursor had to be moved to the appropriate locations to build up 
the computer command to execute the second program, and it 
then had to get back to its original location. After the whole 
process, the screen should look exactly the same as before the 
attack. It was tricky, but doable. Even sitting in front of the termi- 
nal, the victim would only see something similar to a power flicker 
with the cursor flashing around for about a second. The programs 
were not ready until the evening of April 1, and we let them loose 
for the remaining few hours of the day. 

The next morning we saw several unusual messages showing 
up on the electronic bulletin board, with most of the victims being 
faculty members 6 . An interesting message had a faculty member 
at the time, asking who was the owner of a pair of panty hose that 
he found in the ladies' room. Another faculty member "posted" a 
message saying that he managed to log on to the computer, but did 
not know how to log off and needed help. He was not amused, and 
not realizing what day it was, he posted, for real, a message 
demanding that whoever posted the logging-off message confess. I 
resurrected my programs, but this time with only one allowed vic- 
tim, namely me, and the message file was modified to contain only 
one message. The message first pointed out that professors were 

6 Students working late tended to be hackers. Hackers were more likely to disable 
the writing privilege to their terminals. 



An Office of Troublemakers 1 7 

sometimes forgetful and that the previous day was April 1. 
Furthermore, there were clear indications that the messages were 
not real. All of the message titles were enclosed in quotation marks 
and, technically, the message was indeed posted from the profes- 
sor's own account, just as my message was posted from my account 
through the same mechanism. 

I suffered no direct consequence from this incident. But this, 
and my next far more serious prank, probably affected my stand- 
ing in the eyes of some faculty members. 

After this episode, my colleague disabled the writing privilege 
to his terminal permanently. Nobody was ever going to do the 
trick to him! Another friend got interested in playing with the cur- 
sor control and eventually created a "psychedelic command 
prompt". Every time he typed in a command a cowboy-like figure 
would be displayed on his terminal screen, and an arrow would 
shoot out from the left side of the screen and hit the cowboy who 
then disintegrated into nothing. 

Mister Rogers' Neighborhood 

UNIX is a relatively secure operating system. Ken Thompson, one 
of its creators, mentioned several ways to compromise it in his 
Turing Award Lecture. But generally, a well-maintained UNIX sys- 
tem is almost immune to amateurish attacks. The problem is that 
most UNIX systems are not necessarily well maintained. 

In the early 1980s, almost all of the newer computers in uni- 
versities ran UNIX. Many of these systems were connected to the 
DARPA (Defense Advanced Research Project Agency) Net, the pre- 
cursor to today's Internet. The DARPA net was frequently 
shortened to the Arpanet. The CS Department, with a lot of 
research projects sponsored by DARPA, was on the Arpanet. 

The Department had a dedicated and capable staff maintain- 
ing the computer systems, and I never could figure out a way to 
break the main departmental machines. Out there on the Arpanet, 
however, there were lots of less secure machines. 

In the Department, all computer accounts accessible from 
the Arpanet had passwords that had to be entered before logging 
on. This was not true for many computers on the Net. I stum- 
bled 7 onto a machine at UC Berkeley that had a guest account 

7 Well, not precisely. I had friends studying at Berkeley. Another thing was that 1 
was curious about some of the peculiar bugs in the computer game Rogue, which 
was authored by students at Berkeley. 



18 Chapter 2 

without any password. Someone left the door ajar. Now this act 
alone does not necessarily mean there was a problem. The UNIX 
operating system allows you to lock up your own private proper- 
ty so that other users cannot access your on-line love letters. 
(Although why you would leave your love letters on line is 
beyond me.) The guest account, by definition, had limited rights, 
therefore a guest should not be able to unlock the locks. So every- 
thing's OK, right? 

In real life, the building managers usually have master keys 
that they can use to open any door in the building. On a UNIX 
system, the system manager can become a so-called "superuser" 
who can do anything to the system. Certain trusted programs that 
are needed to maintain the system are usually also given the supe- 
ruser privilege so that the programs can function properly. 
Occasionally, the trust placed in these superuser programs is mis- 
placed. The programs themselves may be OK, but the files that 
they rely on to function might be insecure. Any competent sys- 
tem manager would make sure that the superuser programs 
themselves cannot be compromised, but the files referenced by 
the superuser programs can, and do, get neglected by less experi- 
enced system managers. On the Berkeley machine, for whatever 
reason, some of the files were not properly secured, and there was 
a backdoor where a guest could get in and help himself to a mas- 
ter key. 

So I became the superuser, a virtual god, on the Berkeley 
machine. What should I do with the power? Obviously, 1 did not 
want to do anything damaging. A friend made the suggestion that 
a computer account be created for Fred Rogers, the preacher and 
host of the Public Broadcasting System's "Mister Rogers' 
Neighborhood" TV show for children. Fred Rogers lives in 
Pittsburgh, and the show itself is recorded in Pittsburgh, within 
walking distance from Carnegie Mellon. The suggestion seemed 
very appropriate. Soon Mister Rogers got possibly his first comput- 
er account on the Berkeley machine without realizing it. We made 
one mistake in naming Mister Rogers' account: my friend thought 
Fred's middle initial was S, and the account name was given as fsr. 
Mister Rogers' middle name is McFeely. The account name should 
have been fmr. 

One of the regular users on the Berkeley machine was appar- 
ently working on a network project for DARPA, and he had a file 
that contained passwords for machines all over the Arpanet. The 



An Office of Troublemakers 19 

file was protected, but not against a superuser. The passwords were 
also in plain text. Suddenly, I had keys to a lot of normally locked 
front doors all over the Arpanet. 

For the next two weeks, during my free time, Mister Rogers' 
Neighborhood started to expand over the Arpanet. During this 
period, I was constantly hearing in my mind Mister Rogers' song 
from the show, "It is a beautiful day in this neighborhood. A beau- 
tiful day for a neighbor. Would you be mine? Could you be 
mine? ..." 

Creating new accounts for Mister Rogers was slow going as 
each system had a different weakness. I also did not have a lot of 
free time — it was close to the finals for three of the courses that 1 
was taking. But at the peak, Mister Rogers had almost ten comput- 
er accounts at top US universities. 

Then I received a summons from our Department Head, Nico 
Habermann. 

The system manager at Purdue University had noticed an 
account created for Fred Rogers and he followed the trail back to 
Carnegie Mellon. The staff at Carnegie Mellon then found me out. 

Nico gave me a short lecture and I was ordered to write what 
amounted to a confession about how it was done, and to show my 
remorse. It was a slap on the hand. There was no real damage, and 
possibly the systems involved became safer as a result. Mister 
Rogers, of course, lost all his accounts. This was before Congress 
passed the law about computer privacy. I would be in deep trouble 
if my misdeed were done today. It is also unethical to look at pro- 
tected files without permission, and even unprotected files of a 
personal nature should not be looked at without permission. I was 
treating it as a game when in fact it was a serious matter. On the 
other hand, one should never leave a plain text file containing 
passwords lying around, protected or not, and definitely, definite- 
ly, not on a system with guest accounts that do not require 
passwords. The world is a big place, and there are some bad people 
out there. 

I must have been dangerously close to being shown the way 
out in the Black Friday meeting that year. As it was, I was probably 
one more incident away from being asked to leave the 
Department. 



CHAPTER 3 



Taking the Plunge 



Japanese Work Ethic 

I can work hard when I choose to, but basically I am a lazy person. 
By the time I was in my second year at Carnegie Mellon, I had only 
worked hard once for an extended period of time, and that was to 
prepare for my College Entrance Exam in Taiwan '. But studying 
for exams cannot really be considered work. In that sense, until my 
second year at Carnegie Mellon, I had never worked hard in my 
life. Andreas Nowatzyk, my German office mate, was a hard work- 
er, but he worked mostly at night and his work attitude did not rub 
off on me. This changed when I started to work with a Japanese 
visitor on one of Professor Kung's projects. 

Teiji Nishizawa was a visitor from Matsushita Electric in 
Japan. (Matsushita is probably better known in the US as the com- 
pany behind the Panasonic brand of consumer electronic 
products.) At that time the Computer Science Department had 
deals with big Japanese companies that allowed them, for a fee, to 
send representatives to work temporarily at the Department under 
the professors. Teiji was one of these representatives. 

In my experience, there are two types of hard worker. The first 
tends to work in short bursts. Many software programmers fall into 
this type, although not all of the good programmers do. Hard 
workers of this type might work for, say, 48 hours in a row while 
their creative juices are flowing. I can never work in this way; I get 
cranky if I don't have enough sleep. Teiji was a good example of 
the second type of hard worker. He was highly motivated and self- 

l For a discussion of the Taiwanese educational systems, see An Early Riser in A Lad 
from Taiwan, Appendix A. 

20 



Taking the Plunge 21 

disciplined. He might not work long hours, but he would put in 
solid days of work for long periods of time. This type of hard work- 
er is more suitable for chip design work, where being able to 
provide sustained, high-quality output is more important than 
sudden flashes of brilliance. Working closely with Teiji for the next 
year or so turned out to be quite an educational experience for me, 
especially in terms of work ethic. 

Teiji, Alan Sussman (another of Professor Kung's advisees), 
and I worked together on Professor Kung's joint project with 
General Electric (GE). For both Alan and me, this project counted 
as the area qualifier, a required element in getting our PhD degrees. 
The area qualifier was usually some task given by the faculty advi- 
sor to the students to demonstrate adequate grounding in the 
research area. In reality, it was also a way for the faculty to extract 
useful work from the students. I got my first contact with the 
Japanese work ethic during this project. 

Our project was to design a chip to act as the programmable 
communication channels between several of the commercially- 
available floating-point arithmetic chips 2 . One of the possible 
applications was computer tomography for medical diagnosis, 
which happened to be important business for GE. In the early 
phase of the project, Kung and the three of us had weekly meet- 
ings to go over the design. We were still at the stage of specifying 
what the chip would do when Teiji came to a meeting with hand- 
drawn, complete top-level schematics, along with the basic timing 
information. Having never seen anyone working this way, I made 
a personal note, "Wow, this guy works fast." 

GE was responsible for the back end of the chip design, which 
included the physical layout and the chip fabrication, while 
Carnegie Mellon was responsible for the logic design. Teiji, Alan 
and I split the logic design into three sections, with each one of us 

2 There are two types of arithmetic used in computers: fixed point and floating 
point. The simplest example of fixed-point arithmetic is integer arithmetic, 
where the fixed binary point is after the least significant bit (binary digit) of the 
integer. Floating point arithmetic uses two numbers, magnitude and mantissa, to 
represent a single quantity. The magnitude part holds the significant bits; the 
mantissa is a number indicating where the binary point is located and can be 
viewed as a scaling factor. Floating point arithmetic is typically used in engi- 
neering applications. In the early 198ts, it became possible to do floating-point 
arithmetic on a single chip by itself. Today, the floating-point arithmetic unit is 
only a small part of a general-purpose microprocessor. 



22 Chapter 3 

responsible for a section. In the beginning, only Teiji was putting 
in the hours, but both Alan and I soon caught the spirit and 
learned from his example. Andreas Nowatzyk actually helped quite 
a bit as well. We were using a new set of Computer Aided 
Engineering (CAE) tools, of which Andreas happened to be the first 
user. 

The Carnegie Mellon portion of the joint project was com- 
pleted in early 1985. At this point, 1 had finished all the other 
requirements for the PhD degree, with only the small matters of 
proposing, researching, writing, and defending my thesis left to 
do. 

The joint project had taken over a year by then. I could not 
use it as a thesis topic, since there was no significant research con- 
tent. Still, the project experience had made me better disciplined 
and, for the first time in my life, 1 had gone through the entire 
logic design process for a chip. Without going through the project 
and without the subtle influence of our Japanese visitor, I don't 
think I would be ready for my next step. 

Brownie Points 

In 1984 or thereabouts, the Japanese office equipment manufac- 
turer Canon announced a low cost laser printer engine that 
potentially could be used to produce laser printers priced under 
$3000. Within a year, both Hewlett Packard and Apple announced 
the first personal laser printers based on the Canon engine, and a 
new market segment was created. Before Canon's announcement, 
laser printers were mostly used in large departments, and cost 
more than $10,000. The personal laser printer business turned out 
to be very important for both Hewlett Packard and Apple. For 
Hewlett Packard, this development was the beginning of their 
highly successful computer printer business. For Apple, the per- 
sonal laser printer provided the killer application — desktop 
publishing — that allowed Apple's Macintosh computer to pene- 
trate the business market. 

With the proliferation of these new laser printers, it became 
increasingly important to be able to send electronic documents 
from one user to another. To communicate with another user, the 
format of the electronic documents would have to be standardized 
in some way. Xerox pioneered the laser printer business and had 
what was close to an existing standard, but the company was slow 
in getting its standard adopted widely. Adobe, a small software 



Taking the Plunge 23 

startup in California, seized the opportunity and quickly made its 
own new page description language, Postscript, the accepted stan- 
dard. Today, Adobe has a thriving business and a market 
capitalization greater than Xerox itself 3 . 

I had been following the printer business since I was an 
undergraduate, and I found the chain of events started by Canon's 
announcement quite interesting. Canon's engine could print up to 
eight pages per minute, but the printer controller in use in 1985 
had problems printing at the maximum speed even for simple 
pages. If one wanted to print complicated Chinese characters at 
high speed, one would be out of luck. Furthermore, neither Xerox's 
nor Adobe's standard handled Chinese characters. There was a 
window of opportunity to produce a custom VLSI printer con- 
troller that could print more complicated pages at high speed. 
There was also an opportunity to either create a new standard or 
extend one of the two potential standards to handle Chinese, 
Japanese, and other East Asian languages. The business opportuni- 
ties seemed to be there and the ideas probably would make a 
decent PhD thesis topic. I told Professor Kung about my ideas and, 
in early 1985, I started to collect literature on printers, computer 
fonts, and so on. 

It was hard to collect the relevant Chinese literature in the 
United States. However, a paper on the chip we designed with GE 
was accepted for a conference in Taiwan in May 1985, and I had a 
free 4 trip back to Taiwan to collect Chinese literature. Everything 
was looking good. 

Then something unexpected happened. Dr Hans Berliner, 
who was an AI faculty member, asked me to help with the design 
of Hitech, a chess machine that he and a group of graduate stu- 
dents were building. 

I had learned about the Hitech chess-move generator, which 
was used to generate the chess moves directly in hardware for 
Hitech, at the very first seminar that 1 attended at Carnegie 
Mellon. The seminar was a VLSI group seminar; Carl Ebeling was 
the speaker. The move generator was a 64-chip design, one chip 
per square of the 8x8 chessboard. It was neat stuff, but even on first 



3 As of July 2*00, Adobe is worth over $16 billion, while Xerox is worth just over 
$13 billion. 

4 Being a poor graduate student at the time, I found the "free" part important. 



24 Chapter 3 

glance, there was a lingering question at the back of my mind. Was 
it really necessary to use 64 chips? 

At the time that Dr Berliner approached me, an early version 
of Hitech had just been designed and built. I was not the first grad- 
uate student with chip design experience that Berliner had 
approached. Well before I arrived at Carnegie Mellon, Berliner had 
persuaded Carl Ebeling to design the Hitech move generator. Carl 
later became the chief architect of Hitech after agreeing to build 
the whole chess machine. 

A chess machine has three main components: the move gener- 
ator which finds the chess moves, the evaluation function which 
assesses the quality of the positions reached when the chess 
machine looks ahead, and the search control which guides analysis 
of move sequences examined by the chess machine. Typical chess 
programs have the same basic structure, except that the move gen- 
erator, the evaluation function, and the search control are all in 
software. This is also the case for commercial "chess machines" 
that you can buy in a store, as they are really just chess programs 
encapsulated in a stand-alone microcomputer. When a computer 
chess researcher talks about a chess machine, they are usually talk- 
ing about a machine that has at least a hardware move generator 
and a hardware evaluation function. You can expect a hardware 
chess machine to be at least a hundred times faster than chess pro- 
grams running on microprocessors manufactured in the same 
semiconductor technology. When Dr Berliner talked to me, Hitech 
had the 64-chip move generator and a very simple-minded hard- 
ware evaluation function. 

Dr Berliner wanted to improve Hitech's evaluation function. 
He asked me whether I would be interested in designing a 64-chip 
system to compute how well the squares of the chessboard were 
controlled by the chess pieces. I told Dr Berliner that I had a thesis 
topic that I wanted to work on, but if his problem was something 
that could be done in a summer, I might be willing to look into it. 
1 did not mind doing Dr Berliner a favor, but I did not want to give 
him the impression that he could just grab hold of my time. I have 
to admit that I was also looking forward to the chance of working 
with Carl who was putting together a real system, and I had never 
done anything like that. And, of course, the whole idea seemed like 
a good temporary escape from thesis work. 

I had a meeting with Dr Berliner to find out what he had in 
mind. Near the end of the conversation, he mentioned that I 



Taking the Plunge 25 

would not be getting anything more than "brownie points" from 
this work. That was a new term for me. Berliner explained that it 
meant I would not get anything tangible other than appreciation 
for the work done. I was not expecting anything to begin with, and 
that should have been quite clear from the fact that I was only will- 
ing to work on it for the summer. The conversation gave me an 
odd feeling. 

I spent about a week on Dr Berliner's project, entering the top 
few levels of the circuit schematics. I then had a meeting with the 
Hitech group and presented some of my findings. The good news 
was "yes, it could be done". The bad news was "the 64-chip parti- 
tioning makes it necessary to add a lot of external circuits and the 
system would be quite slow". A year later, I discovered that my def- 
inition of "quite slow" would have been more than adequately fast 
for Hitech, but nobody realized this at the time of the meeting. Dr 
Berliner seemed unhappy and asked me to fit the external circuits 
onto the chip anyway. I explained that the system would get even 
slower as the chip input/output delays went up when the slower 
on-chip FET (Field Effect Transistor) drivers replaced the faster off- 
chip bipolar drivers 5 . The meeting was adjourned without any 
decision. 

1 was not a happy camper. The root of the problems appeared 
to be the 64-chip partitioning. The question at the back of my 
mind when I first heard about the Hitech project came back. Was 
the partitioning necessary? 

In the early 1980s, Ken Thompson and Joe Condon, both 
from Bell Laboratories, built Belle, a special purpose chess 
machine, from small and medium scale integrated circuits. Belle 
became the first chess automaton to play at the US National Master 
level in 1983. It had replaced the Chess 4.X as the new top chess 
automaton while I was still carrying out my military service in 
Taiwan. I only learned of Belle's existence after arriving at Carnegie 
Mellon. 

They had written a paper about Belle, describing its basic 
architecture, and I had a reprint at home. That night, after the 
meeting with the Hitech group, I picked up the reprint and reread 



5 Drivers in \C use relatively large transistors, usually of two types: bipolar and 
FET. Bipolar transistors are generally faster due to their higher current driving 
capability. To fully explain what bipolar and FF.T transistors are would take up 
too much space. 



26 Chapter 3 

it. Yes, the move generator described in the Belle paper indeed 
could not be fitted into a single chip. But was it possible to modi- 
fy the design so that it would fit? 

The way I saw it, there were two major problems. 

The first was that the Belle chess move generator had too 
many transistors. The circuit size was simply too big. 

In a chess program, the move generator literally generates the 
chess moves that the program examines. But to work properly, the 
move generator needs to do more than just generate moves. It also 
needs to only generate unexamined, or not yet searched, moves by 
the program. In the Belle design, this second task— of generating 
only unexamined moves — is the job for the clisable-stack. The dis- 
able-stack in Belle is a 64-bit 6 wide memory, one bit for each 
square of the chessboard. 

The disable-stack has two entries for every move searched. A 
normal chess move has a from-square (attacker) and a to-square 
(victim). Each entry on the disable-stack corresponds to one of the 
two types of squares. 

Move generation in Belle has two phases. In the first phase 
(find-victim), a new to-square (victim) is generated. For instance, 
the first phase might generate "rook on h8" as the victim. The first 
entry on the disable-stack is used to mask off all the to-squares 
already searched for this phase. If "rook on a8" is a victim already 
searched, then the a8 square is masked off. 

After a victim has been found, Belle enters the second phase 
of move generation. In this second phase (find-attacker), a new 
from-square (or attacker) is found for the given victim. Say we 
might find that "knight on g6" is an attacker for "rook on h8". The 
second entry on the disable-stack is used to mask off all the from- 
squares (attackers) that are already searched for the given victim. If 
no attacker is left, then Belle returns to the first phase to find new 
victims. 

In all the chess machines that I later designed, I allowed the 
program to look up to 128 plies 7 ahead. With two entries per 
move, the disable-stack would need to be 256 words deep to han- 

6 A bit is a binary digit, either 1 or 0. 

7 "Ply" is computer chess jargon. It means a move made by one of the two play- 
ers. The word "move" might have been more appropriate were it not for the fact 
that in chess literature a "move" could mean a pair of plies, with both White and 
Black making one move each. 



Taking the Plunge 27 

die up to 128 plies of looking ahead. Let us assume that we need 
six transistors for every bit of memory. The number of transistors 
used by the disable-stack alone would be at least 1500 for each 
square of the chessboard, or about 100,000 transistors for the 
whole board. Not a lot by today's standards. But back in 1985, after 
adding in all the other circuits, it was nearly impossible to fit the 
Belle move generator into a single chip. 

What possible ways were there to reduce the number of tran- 
sistors? I had had the reprint of the Belle paper in my apartment 
for more than two years, but had never thought about this ques- 
tion for two reasons. One was that I did not believe that a solution 
to the Computer Chess Problem was within reach. Therefore, I had 
no motivation to probe further. The second was that I had made a 
wrong assumption. I assumed that the Belle designers, as well as 
the Hitech group, must have explored ways to reduce the number 
of transistors. I forgot that the Belle designers, able though they 
were, were not trying to fit their design onto a single chip. They 
designed Belle with discrete off-the-shelf chips, and chips used for 
the disable-stack were fairly easy to come by. For them, the disable- 
stack was perhaps the most elegant solution. It just was not the 
best solution for a single chip design. The fact that the Hitech 
group had adopted the 64-chip design for their chess move gener- 
ator also clouded my judgment. If the Hitech group needed 64 
chips to complete their move generator, how likely was it for any- 
one to create a single chip chess move generator? 

There I was, lying on my bed, thinking about the question of 
the disable-stack seriously for the first time. Once I started think- 
ing about the question, I realized there might indeed be a solution. 
The Belle designers were not looking for an alternative solution 
and Dr Berliner was probably quite happy that the Hitech design 
was different from the Belle design. When Carl Ebeling took the 
job of creating the Hitech move generator, the Hitech group was 
already set with the idea of using 64 chips, one chip per square of 
the chessboard. Consequently, Carl concentrated all his efforts on 
perfecting the 64-chip design. So, even though all the people 
involved were highly capable, they did not have strong reasons to 
look for alternatives. There may very well be an alternative to the 
disable-stack! 

The most radical answer would be simply to get rid of the dis- 
able-stack from the move generator chip completely. The easiest 
way to do this was to take the disable-stack out of a move genera- 



28 Chapter 3 

tor chip and connect 64 wires, one wire for each square of the 
chessboard, from the move generator chip to the external disable- 
stack. It was a possible solution but I did not like it. Ideally, I 
preferred to keep the number of pins 8 for any chip below 40. 
Higher pin-count packages back in those days could cost over 
twenty-five dollars. At that price, the package might cost more 
than the chip itself. A more personal reason for wanting to keep 
the pin count under forty was to see how far I could push the idea. 
The Hitech move generator chips were using 40-pin packages. It 
would be quite a coup if not only a one-chip design was possible, 
but also the new design could use the same package used by the 
Hitech chips. Another reason why I did not like the external dis- 
able-stack solution was that it was really a nine-chip design. Eight 
memory chips would be needed for the 64-bit wide external dis- 
able-stack. The widest memory chip in those days was only 8-bit 
wide. 

At this point, an idea hit me. The new idea rid the wholr 
design of the disable-stack, not just the chip itself. 

The idea hinged on what information the disable-stack was 
really carrying. The disable-stack was keeping track of the history 
of the moves searched so far. In a chess machine, the "move stack" 
contains the very same information, albeit in a different form. The 
64-bit patterns stored in the disable-stack could, in theory, be com- 
puted from the last move searched and currently stored on the 
move stack. The from-square and the to-square of the last moving 
piece contain all the information needed to compute the 64-bit 
disable-mask. I sat up, grabbed a pen and paper, and started work- 
ing out the details. Every one of the sixty-four squares of the 
chessboard is assigned a unique priority number. For a given from- 
square or to-square, we disable all squares with higher square 
priorities. We can compute the 64-bit disable-mask from the 6-bit 
square number by using a six to sixty-four priority decoder' cir- 
cuit. The priority decoder circuit used about ten transistors on 



8 



Silicon chips usually communicate with the outside world through the pins of 
their packages. With some of the newer IC packages, pads might be used instead. 

Here is a simple explanation of a priority decoder. A six to sixty-four priority 
decoder computes a function with a 6-bit input and a 64-bit output. The 6-bit 
input N can be viewed as an integer from to 63. If the 64 output bits are num- 
bered from to 63, then all the output bits numbered below N should be 0s and 
all the other output bits should be is. 



Taking the Plunge 29 

average for each square instead of the fifteen-hundred transistors 
used by the disable-stack. This was a 150 to 1 reduction. How sig- 
nificant was this reduction? My final implementation of the 
complete move generator used about five-hundred and fifty tran- 
sistors on average for each square. Without the design change, the 
number would have been over two-thousand transistors per 
square, and the chip would have been too large by a factor of four. 

I was happily stunned. But there was a second major problem 
with the Belle design. This problem may come as a surprise to a 
layperson. It certainly would have come as a surprise to a much 
younger me. When 1 was a little kid, higher quality radios were the 
ones with more vacuum tubes or more transistors. What could be 
more important to a silicon chip than the number of transistors it 
has? The answer is the number of wires, in particular, the number 
of long wires. The Belle move generator had too many long wires. 

The Belle move generator produces six piece-priority signals, 
one per piece type, for each of the sixty-four squares of the chess- 
board. The piece-priority signals are used to order the moves by the 
values of the pieces, before the simple move ordering by piece loca- 
tions based on the disable-stack. In general, we should firstly search 
moves that capture the most valuable victim with the least valu- 
able attacker. This value-based move ordering can improve the 
efficiency of the chess machine dramatically. 

The 384 (6x64) piece-priority signals get fed into a priority 
encoder, which is a voting circuit used to select the best move. 
The problem lies with the 384 piece-priority signals. They are 384 
long wires. 

For the 3-micron CMOS technology that was available, the 
384 long wires alone would have covered half of the chip. I need- 
ed an alternative way to do the move voting. It took me less than 
one minute to find a working alternative. It was all a matter of 
luck. Back in those days, I subscribed to a lot of IEEE (Institute of 
Electrical and Electronic Engineers) journals. (IEEE gave students 
hefty discounts for journal subscriptions.) A few weeks before that 
fateful night, I had read an article in IEEE Computer Magazine about 
a new computer bus standard called Future Bus. The Future Bus 
used a "distributed arbiter" circuit to vote among circuit boards 
residing on the bus. The "distributed arbiter" required few wires 
and might be a good fit for my problem with the long wires in the 
Belle move generator. The magazine was lying right next to me. A 
quick check of the article verified that a slightly modified distrib- 



30 Chapter 3 

uted arbiter would do the job. Instead of 384 long wires, the num- 
ber of long wires would be reduced to forty-eight, a reduction by a 
factor of eight. 

So, it might very well be possible to fit a modified Belle move 
generator onto a single chip using only a 40-pin package. As a mat- 
ter of fact, there were reasons to believe that such a move generator 
could be more than ten times faster than the 64-chip move gener- 
ator used in Hitech as well. The real chip in the end was closer to 
a factor of twenty times faster. It was over a thousand times 
(64x20) more cost effective. 

Things were suddenly very interesting. Up to this point, I had 
always assumed that the Computer Chess Problem was out of 
reach. My belief seemed to have been confirmed by the 64-chip 
Hitech move generator design. If it took sixty-four VLSI chips to 
complete a relatively slow move generator, then the problem was 
indeed beyond reach. But now it appeared that you could build a 
move generator with only one chip. The Computer Chess Problem 
deserved a second look. 

In the Belle chess machine, besides the move generator, the 
evaluation function was the bulk of the rest of the machine. Could 
the evaluation function be fitted onto a single chip as well? A good 
evaluation function will be far more complicated than the move 
generator. If it was barely possible to fit the move generator onto a 
single chip, what was the chance of doing the same for a good 
evaluation function? It doesn't look good, does it? There was one 
option: namely, we could trade space for time. Chess evaluation 
functions have spatially repetitive components, and it is possible 
to use a smaller circuit multiple times to do the same computation, 
albeit with a time penalty. An interesting fact is that we do not 
have to pay the time penalty all the time. The evaluation function 
can be partitioned into two components: material (basically the 
values of the pieces left on the board) and positional (anything 
else). The material evaluation can be computed very fast. The posi- 
tional evaluation is the part that is time consuming. Usually, the 
positional evaluation also has relatively small values, frequently 
smaller than the value of a pawn. If the material balance is too far 
off, say one side is down a queen, the positional evaluation func- 
tion can be ignored and does not need to be computed. Statistics 
measured later on real chess machines indicated that the full eval- 
uation is only needed about 10-20% of the time. Multiplexing a 
smaller circuit in time to reduce the chip area appeared to be a 



Taking the Plunge 31 

viable option. A back-of-an-envelope calculation showed me that 
the 64-chip evaluation function did not make sense and at this 
point, I lost interest in helping Dr Berliner with his 64-chip evalu- 
ation function. The Hitech system was sufficiently slow that a 
time-multiplexed single chip evaluation function would still be 
more than fast enough. 

The next day I sent an e-mail to Dr Berliner telling him that I 
was no longer interested in building the 64-chip evaluation func- 
tion, and that it was possible to build single chip chess move 
generators and single chip evaluation functions. I also offered to 
do a simple version of the single chip evaluation function, if he 
was interested. Dr Berliner's first reaction was to try to get me to 
"design the whole chip the right way". That is, to design the chip 
as he specified, using the 64-chip design. I made it clear that I did- 
n't want to do the 64-chip design. After a week of going back and 
forth, I sent Dr Berliner a one-page preliminary spec of a simple 
time multiplexed circuit that calculated a subset of the evaluation 
function. Dr Berliner was not entirely satisfied with what I was pro- 
posing. I decided to stand my ground and say that I did not want 
to be involved if my proposal was not acceptable. Perhaps I was a 
little too blunt. I was not planning to do anything more with com- 
puter chess at this point, and I had a serious thesis proposal to 
work on. 

Shortly after, I had a meeting with Professor Kung and told 
him about what had happened. Professor Kung strongly suggest- 
ed that I write up the ideas for the single chip designs and give 
a presentation to a few faculty members. Professor Kung did not 
say it, but from his tone, I got the feeling that I might have 
major problems at the faculty Black Friday meeting (the faculty 
meeting held at the end of each semester to decide the fate of 
each and every graduate student in the department). In my 
blunt response to Dr Berliner I had effectively claimed that 
Hitech was passe, without presenting the reasoning. My stand- 
ing in the eyes of the faculty was already shaky from my earlier 
escapade over the Arpanet. If I wrote the paper and did the pres- 
entation, then the faculty would know that I did not make the 
claim frivolously. For the next few weeks, I dropped my work on 
my thesis proposal and wrote up a technical report. I also gave a 
presentation to faculty members the day before I was to get on 
the plane to Taiwan. 

I spent most of May 1985 in Taiwan. Part of the time was 



32 Chapter 3 

spent collecting books about Chinese fonts, calligraphy and so 
on for the printer controller idea. The rest of the time was most- 
ly just vacationing. I was debating with myself whether I wanted 
to go into the computer chess arena. I did not go to Carnegie 
Mellon to work on computer games; working on computer chess 
would be quite a deviation from my original interest. Moreover, 
the commitment could easily take five to ten years. Writing the 
technical report and doing the presentation, however, made me 
realize that I had the basic blueprint to build the Mother of all 
Chess Machines, a machine that could defeat the World 
Champion. In other words, I had a chance to pursue one of the 
oldest holy grails in computer science, and possibly make histo- 
ry. On the negative side, the printer controller looked quite 
promising as well, and it was not an idea that could wait. 
Personal laser printers were still in their infancy and there was a 
limited time window to make a custom printer controller chip 
widely adopted before the industry standard set in. There was 
some chance that the printer controller idea could make me 
financially independent. There was practically no chance that 
solving the Computer Chess Problem would ever provide any 
financial reward commensurate with the time and efforts 
required. If I went after the Computer Chess Problem, I would be 
going for the glory of knowing that I solved the problem. 1 prob- 
ably would not be financially destitute, but there would be major 
personal sacrifices. The opportunity cost would be extremely 
high. I could be spending many of the most productive years of 
my life on a project that had a very slim payoff. 

The decision was a difficult one. In the end, after serious soul 
searching, I decided to go for the glory. You don't get a chance to 
make history every day. 

At Least Two Years 

Sometimes, nothing is more important than being at the right 
place at the right time. 

If I had not been at Carnegie Mellon, I would not have had 
Mike Browne and Andreas Nowatzyk as my office mates. If Mike 
and Andreas had no interest in chess or computer chess, my inter- 
est in computer chess probably would have completely dissipated 
by the time Dr Berliner asked me to help. If 1 had not had both the 
Belle paper and the IEEE magazine in my apartment on that fate- 
ful night, I probably would not have figured out how to fit a chess 



Taking the Plunge 33 

machine onto a single chip. If Dr Berliner had accepted my sug- 
gestion, I probably would have just done what I could and gone on 
with my original research project. If I had been at some other 
Computer Science Department, and, if Professor Kung had not 
been my faculty advisor, I might not have had the freedom as a 
graduate student to start a project on my own. The fact that it was 
1985 when 1 began the project was also important. Five years ear- 
lier, the technology would not have been sufficiently advanced for 
fabricating a single chip chess move generator. Five years later you 
would be able to put the Belle move generator, as it was, onto a sin- 
gle chip, and somebody else would have beaten me to it. It was 
more or less an accident that I ended up on the path that I took. 

When I first became Professor Kung's advisee, many years before 
the troublemaker label, he told me that graduate students were 
the ones with the real power in the Department, as they were the 
ones with a true grasp of the research projects. Professor Kung was 
probably encouraging me to be more assertive, as many Asian stu- 
dents had the problem of not speaking up. 

After my return from Taiwan, I told Professor Kung about my 
decision to go after the Computer Chess Problem. Professor Kung 
was surprised but receptive. He hinted that I might have only lim- 
ited time to work on the chip. He did not mention the reason, but 
it did not take a genius to realize that I could not afford to take too 
long. The Department already had a computer chess project. 
Starting a new one without fast progress would be an academic sui- 
cide. 

Before I could work full time on the chess chip, something 
else happened. A few days after my return, Professor Kung 
informed me that the GE people were having problems simulating 
the chip that we designed with them and they needed help. 

By this time Teiji had already returned to Japan, and Alan 
Sussman and I were the only people who might be able to help. 
Both of us wanted to spend time on our own research, and we did 
not sign up to do the simulation to begin with. Eventually, we 
both got involved. I took over the bulk of the work in the end, as 
I realized that it was not a waste of time, at least not in my case. 
The chip we designed with GE had quite a bit of transistor level 
circuitry, so we had to do a significant amount of "switch level" 
simulation, where the transistors were simulated as if they were 
on-off switches. The new move generator had not been designed 



34 Chapter 3 

yet, but I knew that it would be mostly dynamic 10 transistor level 
circuitry, and the experience from the GE chip would be quite 
useful. 

The GE work was not finished until late August 1985. Just 
before its completion, I took over the maintenance of the VLSI 
CAD (Computer Aided Design) tools, as the previous maintainer 
had just become a father. It was in my own interest to takeover the 
work, as I was to become the only user of the tools for a while. I 
was ready to begin my new project. 

The Hitech 64-chip chess move generator designed by Carl 
Ebeling took over three years to complete. I did not have that 
much time to spend on my move generator chip. I planned to 
complete my goal of beating the World Champion in five to ten 
years, and I could not afford to spend three years on just the move 
generator. I decided to go full steam ahead and complete the chip 
in six months. Why the six-month target? Well, Joe Condon and 
Ken Thompson finished the Belle chess machine in six months, 
and if they could build a complete chess machine in that time, I 
should be able to complete a VLSI chess move generator. At the 
time, all the chip projects of any significance at Carnegie Mellon, 
and possibly every other university, took over a year to complete. 
If I pulled it off, it would be a speed record of sorts. 

The first phase of the project, the circuit design, was straight- 
forward, and largely drew from my experience with the GE joint 
project. It was hard work, but there was no major surprise along 
the way. The complete circuit design of the move generator, down 
to the transistor level, was finished in one month. The full chip 
was also simulated in its entirety down to the transistor switch 
level in the same period. 

Next came the hard part. The physical layout — the drawing of 
all the geometry, including the transistors and the wires — was the 
next step. Nowadays, most designers would be using automated 
tools to do the layout. The tools available in 1985 could not do it, 
so to a great extent I had to do the layout manually. But there was 
another problem. I wanted to fit the chip into a 40-pin package 
and the die size for such chips was severely limited. The largest die 
size for chips in production quantities today are around 2 cm (0.79 

10 "Dynamic" has special meanings for chip designers. The states or memories of a 
dynamic circuit are stored as charges on capacitors, which have to be refreshed 
constantly. 



Taking the Plunge 35 

in) on a side. The maximum die that could fit into the MOSIS u 
40-pin package was about 0.7 cm (0.28 in) on one edge, or rough- 
ly twelve percent of the area of today's big chips. My task was to fit 
the design into the limited die area and to do it with the tools at 
hand. There was also the small problem that I had essentially no 
experience with the physical layout step of the chip design. 

In the circuit design phase, I reduced the core logic to about 
550 transistors on average for each of the sixty-four squares of the 
chessboard. Owing to edge effects and the fact that a pawn moves 
in a different way on different ranks, the circuits for the squares 
were not identical, but they did share a lot of common logic. As a 
first step, I picked one of the squares and did a trial layout of the 
square. This took about a month. At the completion of the trial 
layout, there was good news and bad news. The good news was, 
that based on the area used by the trial layout, the chip might fit 
with a lot of work. The bad news was that the trial layout I had for 
one square had the wrong shape. It was too tall. To get the maxi- 
mum chip area, essentially, the chip would have to be a square. 
The maximum dimensions allowed were 6812 by 6912 microns, or 
0.6812 cm by 0.6912 cm. Since a chessboard is 8x8, the circuit for 
each logical square would have to be physically laid out as a square 
as well. 

The next two months were spent correcting the shape of the 
layout and creating variations for all the different squares. The 
number of different squares quadrupled when, in order to reduce 
the chip area, it became necessary to tile the squares to an exact fit 
in both X and Y dimensions. By tiling the squares to an exact fit, I 
could share wires and contacts (connections between wires on an 
IC) between adjacent squares and squeeze out a few microns for 
every pair of adjacent squares. The easiest way to do the tiling was 
to mirror the squares in both X and Y directions. This mirroring 
meant that, even for squares with identical circuits, there were four 
different versions of the layout: northeastern, northwestern, 



11 MOSIS (MOS [Metal-Oxide-Semiconductor] Implementation Service) is a low- 
cost prototyping service for VLSI circuits. Initially, MOSIS served only 
government agencies and universities. Today, MOSIS also provides services to 
commercial firms. While I was at Carnegie Mellon, almost all the chips designed 
at US universities went through MOSIS. The chips were effectively free as gov- 
ernment sources such as DARPA and NSF (National Science Foundation) bore the 
cost. 



36 Chapter 3 

southeastern, and southwestern. The final layout of the chip core, 
which was logically an 8x8 chessboard, looked more like a 4x4 
pattern as a result. The pattern was rather pretty and somewhat 
quilt-like. (When Deep Thought, the chess machine based on this 
chip, became widely known, a poster vendor asked me whether he 
could have the right to produce posters based on the die photo.) 
While the end result might have been aesthetically pleasing, the 
two months of hard work had been hellish for me personally. The 
work was tedious and consumed far more time than I had expect- 
ed. I had to work longer and longer into the night to keep up with 
my schedule. Of course, my weekends had already become work- 
days earlier during the logic design phase. 

It was December by this time, and there was another Black 
Friday coming up. Professor Kung asked me to give a presentation 
to the Hitech group. Dr Berliner started to show some interest 
before the presentation but, again, our interests did not really 
align. I was simply going to build a very fast machine, see what it 
could do, and decide what to do next. Dr Berliner had a complete- 
ly different idea. 

Dr Berliner was a leading authority in the computer chess 
field, and his words carried a lot of weight. He was publicly stating 
that brute force 12 searching, such as I was planning, would not go 
very far and it was time to go back to the old idea of selective 
searching 13 , in particular, to the idea of the B-star (B*) search algo- 
rithm, an algorithm that Dr Berliner had come up with in the 
1970s. Most of computer chess researchers appeared to agree with 
Dr Berliner that brute force searching was reaching its limit, but 
the suggestion that B* could rule the day was not universally 
accepted. 

My intuition was that B* was a dead end. Anyway, I gave my 
presentation and Dr Berliner gave me references about B*. I did not 

12 Brute force searching in computer chess means logically examine every possible 
move, in contrast to selective searching where some of the moves might be 
pruned away. 

1! Selective searching means different things to different people. In general, selec- 
tive searching has two flavors. One is to search "interesting" lines deeper (also 
known as selective deepening). The other is to search "uninteresting" lines shal- 
lower, or not at all, that is, to prune away the "uninteresting" lines (also known 
as selective pruning). It is possible to mix the two flavors in the same program. 
Some people consider selective deepening as modified brute force searching and 
not selective searching. 



Taking the Plunge 37 

really have the time to go over his references until the chip was 
done. 

Andreas Nowatzyk had frequent contact with several mem- 
bers of the Hitech group. A few days after my presentation, he told 
me that the Hitech group, or at least someone in the group, 
believed that it would take me "at least two years" to finish the 
chip design. I was quite amused. I had just given the Hitech group 
a presentation about a chip that was completely simulated, and 
almost completely laid out, and yet at least one person still 
believed that it would take me two more years to complete the 
design. I had completed the layout of all the squares, and I could 
see that I had a good chance to finish the chip in six months as I 
planned. The two-year prediction was probably based on how long 
it had taken Carl to finish the Hitech move generator, which was 
about three years. I was not necessarily a better chip designer than 
Carl, but I had better tools, I did not have to answer to a commit- 
tee who changed the design spec from time to time, and I was 
determined to finish the chip in as short a time as possible. If it 
meant that I had to spend long hours on the project for sustained 
periods, I would not hesitate to do so. On top of all these reasons, 
I had Carl's indirect help. During his work on the Hitech move 
generator, he produced a tool that could be used to compare one 
netlist (the list of all the transistors and the connections between 
the transistors) with another. By comparing the netlist for the lay- 
out with the netlist for the logic design, a designer can verify that 
the layout matches the design. Carl's tool was quite important in 
the final verification phase of the design. 

After two months of laying out all the squares of the chess- 
board, I had a new problem to deal with. Even with all the squares 
fitting tightly together, there was not much space left on the rim 
of the chip. In the remaining space, I had to squeeze in the edge 
wiring for the squares, the control logic for the chip, the interface 
logic to the world outside the chip, the wiring between the inter- 
face logic and the IO (input/output) pads, and, of course, the IO 
pads themselves. The main trouble was with the pads. The exist- 
ing pads available from MOSIS simply would not fit. The pads 
connect the chip, usually through mechanically bonded gold 
wires, to the outside world. Given the mechanical nature of the 
wire bonding process, the wire landing area for an IO pad is fair- 
ly large in comparison to other chip features. The wire landing 
area, however, is only a small portion of an IO pad. The rest of the 



38 Chapter 3 

area is taken up with special protection circuits, which protect the 
chip from external electrostatic discharges, and the IO driver cir- 
cuit, which boosts the internal signal to adequate strength for 
driving external circuits. The MOSIS IO pads were long rectangles, 
with the wire landing area near one of the short edges of the rec- 
tangle. The problem with these pads was that the long edges of 
the pads were perpendicular to the chip edges. They were fine if 
you needed to place, say, more than a hundred IO connections 
around the chip, but not so good if you didn't need a lot of IO 
connections. With the MOSIS pads, the IO ring around the chip 
would be a very thick ring. If I decided to go with the MOSIS pads, 
the move generator chip would have been about 1000 microns 
too wide, and too tall, or roughly thirty percent too large in terms 
of die area. 

I had designed my move generator to use no more than 40 
pins, so I did have an alternative, namely, getting a new set of 
pads. Viewed from the chip edge, the MOSIS pads were tall and 
skinny. My new pads would have to be short and fat. Since there 
was no expert with design experience for IO pads available, I just 
went ahead and designed my own. The layout of the pads and the 
remainder of the chip circuits took about a month. Five months 
from the start, I had a first cut of the complete chip layout. It bare- 
ly fitted into the maximum die area allowed. I had no more than 
five microns to spare in either X or Y direction — not even enough 
to squeeze in one more wire. Talk about cutting it close. 

At this point I was feeling good and getting bold. It would be 
another year before I presented my thesis proposal to the 
Department at large, but I was quite certain what my thesis would 
be. 1 also had a good idea about my choice of the outside thesis 
committee member. At Carnegie Mellon, the thesis committee nor- 
mally had three members from the university itself and one 
outside member. Professor Kung would automatically be the head 
of the thesis committee and normally he would play a leading role 
in finding the outside member. Without telling Professor Kung, I 
sent an e-mail directly to Ken Thompson, asking whether he would 
mind being on my thesis committee. I did not tell Ken the minor 
detail that I did not have a thesis committee yet. Normally, you 
don't just ask a Turing Award winner to be on your committee; I 
was relying on the shock value of the request to get Ken interest- 
ed. He was suitably shocked, and asked for more information on 
my work. I sent him the technical report that I had written before 



Taking the Plunge 39 

my trip to Taiwan and, after reading it, he agreed. I don't know 
why I did not ask Dr Berliner, who was very well known in the 
computer chess field, to be on my committee as well. It might have 
been that I simply wanted to emphasize that the thesis was really 
about hardware design and computation theory instead of com- 
puter chess per se. With hindsight, however, I might also have had 
some subconscious concern about a potential conflict of interest. 

I had one month left of my original six-month schedule. The 
chip still had to be verified. First, I needed to re-simulate to verify 
the chip's functionality. This was straightforward, but there was 
one thing that I had to do before the re-simulation. During the lay- 
out process, I made changes to the circuit to allow more efficient 
layout, but I had not yet updated the circuit schematics to reflect 
the changes. The updates were made and the circuit was re-simu- 
lated. The next step, and this was where Carl's tool came in, was to 
verify that the re-simulated netlist and the physical netlist were 
identical, transistor for transistor and wire for wire. There was a 
slight problem associated with this last step — Carl's tool had its 
own idea on how the netlists should look. I wrote a translator to 
create netlists of the right form, and then compared the two new 
netlists. Usually, most designers would compare the netlists in a 
bottom-up fashion, that is, starting by comparing the netlists for 
the smallest sub-circuits, then the netlists for higher levels of cir- 
cuits, and eventually the netlists for the top level. Just in case I got 
lucky, I compared the top level of the netlists first. 

Not surprisingly, the comparison failed. There were too many 
mismatches. In particular, the power and the ground were shorted 
(directly connected) in the physical netlist. I would have one HOT 
chip that could be used to boil an egg if I submitted the chip the way 
it was. On the top level of the circuits, there were tens of thousands of 
connections to the power and the ground pins. Naturally, it was not 
possible to tell from the top level of the netlists where the power and 
the ground were shorted. The obvious strategy was to divide and con- 
quer, namely, to compare the two netlists in the bottom-up fashion. 
Once the real netlist comparison started, I found numerous other 
errors. Luckily, all the errors could be corrected without increasing the 
already tight die size. By February 1986, six months after I started, the 
chip was designed, laid out, verified and ready for fabrication. 

Now, the big question. Would the chip work the first time? Or 
would the person in the Hitech group be correct in predicting that it 
would take me at least two years to get my move generator working? 



40 Chapter 3 




A Silicon Chessboard. 

Aluminum on silicon canvas, drawn in 7 985-7 986, Feng-hsiung Hsu. 

Also acts a single chip chess move generator and was the core of 

the top chess machines from 1987 to 7 995 



Tom Sawyer's Trick 

The semiconductor industry went through a major transition in 
the mid-1980s. In the early 1980s, NMOS (N-type MOS [Metal- 
Oxide-Semiconductor]) was the low cost, high volume technology. 
By the mid-1980s, CMOS (Complementary MOS) was slowly 
replacing NMOS as the dominant low cost technology. The CMOS 
technology uses both the P-type MOS transistor, where the main 
charge carrier has a positive charge, and the N-type MOS transis- 
tor, where the main charge carrier has a negative charge. CMOS 
circuitry has lower power consumption, but the process is also 
more complicated than NMOS. 

MOSiS, the silicon broker providing chip foundry service to 



Taking the Plunge 41 

US universities, started to offer a CMOS fabrication service in about 

1985. In the beginning, few university researchers used the CMOS 
service. One reason was the lack of experience with CMOS design, 
and a second was that the CMOS process provided by MOSIS was 
not mature yet — rumor had it that someone got working CMOS 
chips from MOSIS, but no one saw them. To be fair to MOSIS, most 
of the designs submitted to MOSIS, either NMOS or CMOS, were 
probably flawed to begin with. 

I submitted the CMOS chess chip in March 1986. Normally, it 
took about six to eight weeks before the packaged chips came back, 
and I had some time to work on something else. I took a closer 
look at the references related to B*, but I was not impressed. The 
reading confirmed my first intuition ,4 . Contrary to the claims, the 
numbers appeared, at least to me, to show B* to be inferior to the 
simple-minded approach of just searching every move to the same 
depth — the so-called brute force approach. 

Of course, some sort of cooperation with Dr Berliner might 
still be possible. After all, I would need a good chess player on 
the team, and Dr Berliner was the best player around, being a 
former World Correspondence Chess Champion. 

Up to this point, I was making the assumption that I would 
find some way to parallelize the search. There were two steps in my 
plan to create a chess machine a thousand times faster than Belle. 
The first was to create a chess machine that was either a single chip 
or a small number of chips. The second was to find a way to use a 
massive number of chess machines in parallel to effectively speed 
up the search process. Given that the single-chip chess move gen- 
erator had already been designed, I was part way through the first 
step. The second step was, however, an entirely different matter. At 
the time, no one knew how to effectively parallelize the search, 
despite many years of active research. I simply believed that a solu- 
tion would be found. It did not matter to me whether I would be 
the first person to find a solution; as long as one was found before 
the final encounter with the World Champion, I would be fine. Of 
course, it would be so much sweeter if I could find the solution 
myself. 

The problem turned out to be simpler than I thought. By April 

1986, I had a decent solution that appeared to work well in pre- 

14 A B* version of Hitech appeared in early 1990s. Its public tournament results 
were no better than those of the brute force version of Hitech. 



42 Chapter 3 

timinary simulations and could be proven theoretically to have 
some fairly good properties. At about this time, I heard that the 
fabrication run for my submitted chip had run into some problems 
and it would take quite a while before the chips arrived. Shocked 
by the news, I submitted the chip again in early May. Two days 
after my submission, MOSIS sent me a message saying that the 
chip I submitted in the first run was too large to fit into the 40-pin 
packages that I specified. I informed them immediately that the 
chip was at exactly the right maximum size. A few hours later, I got 
an apology — a software glitch had added a few microns to my chip 
in their calculation, and the chip got bumped into a larger die. The 
chips would be packaged in 64-pin packages, which were about 
three times bigger than the 40-pin packages. I did not know for 
sure that the second run would go correctly and, given the succes- 
sive glitches, I submitted the chip twice more in late May and late 
June. 

I received the first thirteen chips on 28 June 1986. To my sur- 
prise, they were from the second run and were packaged in 40-pin 
packages. The foundry used on this run had a bad reputation, and 
I was a little bit apprehensive. I measured the resistance between 
power and ground, and several of the chips were shorted. My heart 
sank a little. The CAE workstation that I used to design the chip 
had an attached hardware hookup that could be used to test the 
chips functionally. I hooked up the chips that were not shorted. 
No go. None of the chips showed any glimmer of life — they were 
not passing even the simplest test. I opened some of the chip pack- 
ages and placed the chips under a microscope; defects were plain 
to see. I sent a report to MOSIS, telling them about the defects. 
Their first response was a request for one of the bad chips. Then I 
got another response questioning whether I had worked on any 
large project or had any experience testing chips. I was more than 
a bit annoyed, to say the least. Apparently, no other people were 
reporting zero yields on the very day they received the chips. It was 
very likely that I was the only one complaining at the time. 

The next batch of chips arrived two weeks later on July 14. 
These were from the long-awaited first run, and were packaged in 
64-pin packages. The test setup was for 40-pin packages. Andreas 
soldered together a 40-pin socket with a 64-pin socket to act as an 
adaptor, and I was able to test the new chips using the adaptor on 
the old setup. The chips behaved much better this time. Most of 
the chips passed some tests, and two of the chips passed all the 



Taking the Plunge 43 

tests. Bingo. The chip worked the "first time". The percentage of 
good chips (also known as the yield) was a little bit low, and look- 
ing at the non-working chips under a microscope showed that the 
metal wires were wider than they should be. 

I sent out another report to MOSIS, mentioning that these 
chips were the same design as in the zero-yield run. This time 1 got 
a reply indicating that the chip vendor dry etched (etching without 
using wet chemicals) the metal layer, but MOSIS had prepared the 
mask for wet etch. This explained the initial delay for this run. A 
few days later, the third run arrived, with chips in 40-pin packages. 
This run had better yield and about half of the chips worked. The 
fourth run arrived a month later, was fabricated by the same ven- 
dor as the second run, and again had a zero yield. MOSIS this time 
took the report seriously, put the vendor on the watch list, and 
even suggested using the chess chip as a test chip for qualifying the 
vendor. 

When I got the first two "working" chips, I verified that the 
chips worked as designed. But the design itself was not verified to 
work within a real chess program yet. 

At Carnegie Mellon, the CS Department had a support organ- 
ization known as the Engineering Lab to take care of the 
departmental network and machines, and to provide engineering 
support for the researchers. Several of the staff engineers there were 
good friends of mine. One of them, Lawrence Butcher, lent me a 
computer bus interface card and gave me a small empty wire-wrap 
board. (A wire-wrap board is a prototype board where circuit con- 
nections are made by wrapping wires around posts on the backside 
of the board.) I wire-wrapped a simple nine-chip chess chip inter- 
face on the wire-wrap board. Lawrence helped me make a cable 
connecting the two boards. Now, the problem was finding a 
machine to put the bus interface card in. One possibility was to put 
it in one of the machines in the Engineering Lab, but the Lab was 
on a different floor from my office. Then I hit upon a great idea. 

In Mark Twain's classic novel The Adventures of Tom Sawyer, 
Tom was ordered to paint the fence. He pretended that it was fun 
painting the fence, and got the other kids to think that it was such 
fun that they were willing to pay Tom to do his work for him. 
Being a lazy person myself, this story made quite an impression on 
my young mind when I first read it as a small boy. 

Mike Browne, my office mate, wrote a chess program while I 
was working on the chip. I tried to interest him in porting his pro- 



44 Chapter 3 

gram to run with the chess chip, not that I wanted to have a com- 
petitive chess program, but ] did want to make sure that the chess 
chip worked. If I could get the chip to run inside a working chess 
program, preferably with someone else doing the work, I would be 
sure. My direct approach with Mike did not work. Remembering 
Tom Sawyer, I went for a sneakier approach. 

Thomas Anantharaman and I were in the same incoming 
class of CS graduate students. He was in the speech group, but he 
was also in the VLSI group that I was in, so we had a lot of contact. 
Thomas had quite an unusual upbringing. His father, a retired pro- 
fessor of Metallurgical Engineering, was also an expert Yoga 
practitioner. His German mother was a student of Indian 
Philosophy. Thomas graduated from the Institute of Technology at 
Benares Hindu University in India. I don't know how, but he seems 
to be capable of thinking in different ways than I do; sometimes, 
he reaches conclusions that I would never dream about in a thou- 
sand years. Thomas' office was on the same floor as mine and it 
was only a few doors away. 

Like Mike, Thomas had also written a chess program for his 
own amusement, and he had an office workstation that was com- 
patible with my interface to the chess chip. Had I known how my 
next action would impact on Thomas' thesis research, I would 
have refrained from doing it (although, when all was said and 
done, he did all right in the end). Anyway, I persuaded Thomas to 
let me put the interface card into his workstation. I then wrote a 
more detailed testing program to test the chips. Whenever a new 
batch of chips came in, I could test them all within five minutes, 
at about ten seconds per chip. I placed the source of the testing 
program in a public place that Thomas could access, and then I got 
back to my simulation on how to parallelize the search, pretend- 
ing to ignore Thomas for the moment. 

A few days later, Thomas informed me that he had adapted 
his chess program to use the chess chip, just as I had secretly 
planned. Thomas was able to get his program, with the chess chip, 
to search about 30,000 positions/sec. The fastest PC-based chess 
programs today can search at least 10 times faster, but in 1986 only 
special purpose chess machines, and chess programs running on 
supercomputers, were faster. The first game that Thomas' new pro- 
gram played was against Murray Campbell, a fellow graduate 
student and chess expert, as well as a student of Dr Berliner's. 
Murray had done some work on Hitech, although not as part of his 



Taking the Plunge 45 

main research. He certainly had good ideas about how to play 
chess against computers. Yet, Murray lost the game. I was quite cer- 
tain that the chip was correct after my own detailed testing, but 
seeing it used in a working chess program dispelled any residual 
doubt. 

At the time that Thomas got involved, I had not really 
planned on forming our own computer chess team. I had assumed 
that I would find some common ground and work together with 
the Hitech group. Getting Thomas to port his chess program was 
just a short cut to avoid doing a piece of work that I should be 
doing myself. My plan at the time was to create a three-chip chess 
machine. Besides the already completed move generator chip, the 
two new chips would evaluate chess positions (the evaluation 
chip) and control the search (the search controller). I never did 
complete the chip set because of technical difficulties. One of them 
was that the evaluation chip would need to have RAMs (Random 
Access Memory) on it. Designing RAMs without detailed process 
control, as in the case of MOSIS service, was no trivial task. The 
other major technical difficulty was that I really did not know that 
much about chess at the time. When Thomas and I decided to 
enter into computer chess competitions later on, the plan for the 
chip set became permanently shelved. 



CHAPTER 4 



The Chess Machine That Wasn't 



Seven Weeks 

Thomas Anantharaman's exploit in creating a very fast chess pro- 
gram did not go unnoticed in the Department. The fastest chess 
programs in 1986 were Belie, Hitech and Cray Blitz. Belle and 
Hitech were special purpose chess machines roughly the size of an 
office refrigerator. Cray Blitz was a chess program running on a 
multi-million-dollar Cray supercomputer. The top programs at the 
time were searching in the range of 100,000 chess positions per 
second. Thomas was getting 30,000 chess positions per second 
with a vanilla office workstation plus a chess chip interface card 
about the size of a paperback novel, although with a brain damaged 
chess evaluation. 



The frequent TG (Thank God [It's Friday]) parties were some of the 
unofficial fringe benefits of being a graduate student in the 
Department. They were not always held on Friday, and could be 
held for all sorts of reasons. The Department sponsored some of 
them, but the graduate students and faculty members sponsored at 
least half i. 

It was at one of these TGs that Gordon Goetsch, a graduate 
student working with Dr Berliner on Hitech, jokingly suggested 
that Thomas and I enter the program into the annual ACM 
(Association for Computing Machinery) Computer Chess 
Championship upcoming in Dallas. Gordon was of the opinion 



Later on when our chess machines, ChipTest and Deep Thought, scored major 
tournament victories, we sponsored four TGs ourselves to celebrate the wins. 

46 



The Chess Machine That Wasn't 47 

that the programs in the tournament, despite claims to the con- 
trary, had very little chess knowledge, and a fast and dumb 
program, such as Thomas', would be able to do serious damage. It 
would not win — there were faster programs that were not as 
dumb— but it would be fun to watch what would happen. Neither 
Thomas nor I took Gordon's suggestion seriously at the time. 

After the TG, Andreas Nowatzyk and I ended up in the 
Engineering Lab with some of the staff engineers. I did not intend 
to do what Gordon suggested, but I had a lingering question about 
the chess chip. With Thomas' program running at 30,000 chess 
positions/sec, I knew that the chess chip worked, but I had no idea 
at what speed it could run without overheating, if I stuck with my 
original plan of building a chess chip set, it would be at least 
another year before I knew the answer. That would be a risky 
proposition. I probably should build something that could exercise 
the chess chip at a higher speed. If I built such a speed tester, then 
modifying it into a "poor man's chess machine" and entering the 
ACM competition would be quite interesting. Of course, there were 
a few obstacles. 

First, there were only seven weeks left before the ACM com- 
petition. Second, I only had a student budget, namely, nothing. 
Last but not least, I needed help if I wanted to participate in the 
ACM Tournament, with a competitive program, in seven weeks 
time. 

Andreas provided the first piece of help when I mentioned the 
idea of the speed tester to him and the staff engineers. His eyes lit 
up and he suggested a source for getting a prototype board that 
could be used to make the tester — the departmental scrap heap for 
Perq workstations. The first computer workstation to hit the mar- 
ket, the Perq, was both before its time and beyond its time. The 
workstation market eventually took off, but the Perq, stuck with an 
older multi-chip CPU (central processor unit) design, was taken 
over by nimbler competitors that used cheaper single-chip micro- 
processor CPUs. Three Rivers Computers, the manufacturer of 
Perq, went out of business in 1986, and the Department retired 
most of the Perqs. The scrap heap was full of Perq prototype 
boards, and it would be a simple matter to outfit one of them with 
wire-wrap sockets as a substitution for a full-sized, wire-wrap 
board. Lawrence Butcher, the engineer who lent me the bus inter- 
face card, volunteered to wire wrap the board when the design was 
complete. There was one more problem — getting the electronic 



48 Chapter 4 

parts for the tester. Professor Kung had a lot of spare parts in his 
cabinet for a DARPA project. I decided to borrow some parts from 
the cabinet — if they ever became needed, they could always be 
reclaimed from the tester board. That night, I took an inventory 
list of the spare parts in Kung's cabinet, and Andreas located a Perq 
prototype board in decent shape. The tester would be designed 
using only the parts in Professor Kung's cabinet, although he did 
not know about the borrowed parts until after the chip tester was 
completed. 

The next morning, I had a chat with Thomas. I briefed him 
on the chip tester, what the new software interface would look like, 
and what kind of new hardware functions would be provided. The 
most important new function was the ability to do searches direct- 
ly in the tester hardware, which would then allow us to search 
faster and exercise the chess chip at higher speed. The searches per- 
formed in hardware would necessarily be simple minded, given the 
design time constraint. The hardware searches would not know 
about either castling moves or underpromotion moves (promoting 
a pawn to a rook, a bishop or a knight, instead of a queen as in a 
normal pawn promotion). That is, the hardware would not even 
know all the legal chess moves. Moreover, the hardware would not 
know that repeating a position could lead to a rule draw in chess. 
Also, since it was going to be just a chip tester, the hardware would 
evaluate a chess position based only on a piece-placement table (a 
table tabulating the value of every piece on each square of the 
chessboard). It could be used as a chess machine, but a very crude 
one. To compensate for the hardware limitation, the first few plies 
examined along any line during the search would be done in soft- 
ware, thus knowing about all the legal moves and the repetition 
draws. The last few plies would be done in hardware. The overall 
speed would be close to the hardware speed, since there are far 
more new positions, say, 10 plies away from the root 2 position 
than new positions, say, six plies away. It was also possible to add 
more complicated chess evaluation in the software part of the 
search tree. 

The changes that we needed to make to Thomas' program 



2 The positions traversed by a chess program during a search can be viewed as a 
search tree, with the position on the physical chessboard acting as the root posi- 
tion. The moves searched become the tree branches, and the positions searched 
become the branching points. 



The Chess Machine That Wasn't 49 

were relatively simple, and he agreed to make them. Thomas was a 
good programmer but neither of us was a chess player. We needed 
at least one decent chess player to write the evaluation software 
and to prepare the opening book (the moves that the computer 
would play at the beginning of the game). 

I tried to get Mike Browne interested again. He was one of the 
best programmers that I knew, and he had played chess competi- 
tively in high school but he was preoccupied with other things. 
There were a few other graduate students who were good chess 
players in the Department. One of them, Peter Jansen, expressed 
some interest, but he was busy with course work. 

Murray Campbell, the first victim of Thomas' hardware-assist- 
ed program, was perhaps the strongest chess player among the 
graduate students, and he had worked on computer chess before. 
Murray was a chess prodigy, twice representing the Province of 
Alberta in the Canadian Junior Championship. He had quit play- 
ing serious chess when, in his last chess tournament while he was 
still in the prize hunt, he agreed to an early draw so that he could 
play tennis. In his words, "That was when I knew that I didn't 
want to play competitive chess any more." He would have been a 
great addition to the team, but there was one complication. He was 
in the Hitech group, and it would not be proper for me to go after 
him. 

Two weeks after the conversation in the Engineering Lab, I got 
a surprise e-mail from Murray asking whether the piece-placement 
table on the tester board was adjustable after each move. If it was, 
he had some ideas on how to get the program into reasonable posi- 
tions after the first few moves, assuming that we were interested. 
The table was indeed adjustable and, after finding out that Hitech 
was probably not going to play in the tournament, we took him up 
on his offer. I submitted the entry, without a finished design and 
without a working program, to the ACM tournament committee at 
about this point. We called the new program ChipTest, in part 
because of the chip tester nature of the hardware, and in part 
because we wanted to stress that it was not a real chess machine. 

Four weeks into the seven-week period, the hardware design 
began taking shape. I felt comfortable enough to go out in the 
open and posted an electronic message seeking additional help. 
Since Murray had not promised to work on the evaluation func- 
tion, I talked with a few other chess players but with no luck. 
However, I did get one positive response to my message from Guy 



50 Chapter 4 

Jacobson. Guy was not just a garden-variety graduate student. He 
was one of the authors of Rog-O-Matic, an automatic program that 
played rogue, a popular computer game among graduate students 
at the time. By all accounts, he was one of the premier program- 
mers in the Department. 

Thomas, Murray, Guy, and I had a meeting. Thomas would do 
the bulk of the programming. Murray would take up the tasks of 
preparing the opening book and writing software for the piece- 
placement table, which was effectively the main evaluation 
function. Murray expanded his involvement when it became cer- 
tain that Hitech would not play in the upcoming competition. 
Guy would work on a software evaluation for pawn structure so 
that the program would know how strong or weak the pawns were 
in the software part of the search tree. 

The hardware was ready for wire wrapping by the end of the 
fifth week, and Lawrence Butcher completed overnight the job 
which would have taken me a few days. The hardware had about 
ninety chips which would have cost about $500 if we had had to 
acquire the parts ourselves. 

Because of the time constraint, I designed the tester with a 
minimalist attitude. Unfortunately, I went too far — the tester con- 
tained absolutely no circuitry for testing the tester itself. The tester 
was, however, microprogrammable; that is, it could be pro- 
grammed with a microprogram, which controls all the hardware 
circuits directly. In theory, the tester could be tested to some 
extent, assuming that I had the testing microprograms. But in 
order to verify the correctness of the testing microprograms, I 
needed to have working hardware. It was the classic chicken and 
egg problem. 

Andreas came to my rescue by creating a high quality 
microassembler (a program that converts the microprograms into 
appropriate bit patterns for controlling the hardware) that made the 
coding of the microprograms much easier. It still took about one 
week and several microprograms to complete the testing of the hard- 
ware. There was one more microprogram to write — the one that 
controlled the hardware search function. Time was getting short. 

Meanwhile, Guy did not get around to work on the program 
which meant we would not have a software evaluation for pawn 
structure. Oh, well, Guy's help was a gift to begin with. 

It wasn't time to panic yet, but we needed to have a contin- 
gency plan. We could still play without the hardware search, but 



The Chess Machine That Wasn't 51 

we would suffer a heavy speed penalty. The search microprogram 
was much bigger than all the test microprograms combined, and it 
was also far more difficult to debug. I did not sleep the night before 
I was to leave for Dallas, site of the ACM tournament, but contin- 
ued working until I had to get on the plane. Andreas drove me to 
the airport, and I was silent the entire way. Thomas would pick up 
the work on the microprogram while the competition was going 
on. We did not have a working microprogram for the entire tour- 
nament. 

The previous year I had asked Ken Thompson to be on my 
thesis committee, but I only met him for the first time in Dallas. 
He was there as a tournament official, but after Cray Blitz pulled 
out, citing unavailability of machine time, Belle was drafted to 
play and Ken became one of the competitors. The top seed that 
year was Bebe, and we were paired against the Bebe team in the 
first round. The authors of Bebe were Tony and Linda Scherzer, a 
couple from the Chicago area. My first conversation with Tony 
did not go too well. I tried to explain why ChipTest would not be 
at full speed by telling Tony that we had started to build it seven 
weeks before the tournament, and that we needed another week 
or two before we could get to the top speed. Tony said, "Yeah, they 
all said that they were off by a week." To him, I must have sound- 
ed like a braggart (and there were plenty of them around in the 
computer chess world). Bebe, which was another special purpose 
chess machine, proved to be the stronger player and ChipTest 
went down without much of a fight. Tony turned out to be a nice 
guy when I got to know him better. He was quite open about what 
he was doing and I had many mutually beneficial conversations 
with him. 

In the next round we played Monty Newborn's Ostrich. 
Monty organized the ACM tournaments from their beginnings in 
1970, and this was Ostrich's last tournament. Ostrich was a con- 
tender in earlier days, but it was no longer considered a strong 
program. ChipTest got a winning position fairly easily against 
Ostrich, and then disaster struck. There were still a few bugs in 
both ChipTest's software and hardware, and Thomas put in a 
"cure". Unfortunately, the "cure", being very simple minded, 
turned out to be worse than the disease. Whenever the program 
detected that it was having problems, it would just play the first 
move generated by the hardware move generator, which was effec- 
tively a randomly chosen move. In this game, ChipTest played a 



52 Chapter 4 

move out of the blue as the result of the "cure". The move sacri- 
ficed a rook for a pawn, and with the rook, the game. Thomas 
worked overnight and replaced the "cure" with a safer alternative, 
that is, to record what was previously the best move and whenev- 
er the program ran amok, play the recorded best move. 

ChipTest won the next game against Rex without much diffi- 
culty, and recorded its first win. The next round, it played Merlin 
from Germany, a well-respected program. ChipTest won the game 
by outsearching Merlin, but not without some incidents. It played 
an illegal castling move, but it also stated that the castling move 
was illegal and that the program was in an inconsistent state. The 
tournament director and our opponent agreed to let us restart the 
program. It played a legal move after the restart, and ChipTest 
went on to win the game. 

In the last round, ChipTest played against Recom. Recom was 
a predecessor of Rebel, one of the best commercial chess programs 
today, and it was a strong program even back in 1986. ChipTest 
was playing well early on, but the problems from the Ostrich game 
resurfaced. With the new "cure", it played moves that did not get 
killed outright, but the position was getting really bad. 

Then a miracle happened. ChipTest suddenly claimed that the 
game was drawn, and it played a move that sacrificed another rook 
for a pawn. But unlike the sacrifice in the Ostrich game, this was a 
good sacrifice. I got so excited that I was jumping up and down. 
Recom, after a while, realized that the game was indeed drawn, and 
we called the truce. Every time that I met Recom's operator since, 
he always brought up this scene of my jumping up and down at 
the end of this game. 

For some unknown reason, the wrong game score was posted 
for the Recom game in the official bulletin. Dr Berliner found the 
posted final position a dead loss for ChipTest and was surprised 
enough to e-mail me asking whether Recom had crashed. No one 
else seemed to have caught the mistake. The real game score is 
given in Appendix B, courtesy of Ken Thompson. 

ChipTest's debut was not quite a coming-out party, but an 
even score (two wins, two losses and one draw) on the first try was 
not exactly a bad result. 

Singular Extensions 

Despite being the top seed at the 1986 ACM tournament, Bebe 
did not win the Championship. Given the use of the "Swiss" 



The Chess Machine That Wasn't 53 

tournament format, this was not really a surprise. In a Swiss 
tournament, players with similar scores who have not yet played 
against each other are paired together for the next round. A Swiss 
tournament is better than a knockout tournament in that there 
is some chance to recover from an early loss, but the odds are still 
not good for the top player to win. The probability of the 
strongest program winning the ACM or a similar tournament is 
actually quite slim. 

In 1986, Bebe lost to Belle, the eventual winner. In the battle 
for second place, Bebe was defeated by Lachex (Los Alamos CHess 
Experiment), a program running on a Cray supercomputer at the 
Los Alamos national lab. Tony Scherzer was quite disappointed 
about the loss. He told everyone who would listen what had hap- 
pened in the game, and it was an intriguing story. 

Bebe and Lachex somehow went into a sequence of forced 
moves where both sides had only one single good choice along the 
way. Neither program had any idea about who would come up 
ahead in the end. Both sides just kept on playing the "only" moves 
until, suddenly, both programs realized that Bebe was a dead loss. 
Bebe was one of the top programs, and Lachex was no slouch 
either. How could both programs be so clueless? And what could 
be done about the problem? 

The ACM tournament was not just a competition; it was also 
the annual gathering of computer chess researchers. Between 
rounds, impromptu discussions took place whenever a small group 
of people got together. One of the discussions that I had with Tony 
Scherzer and a few others was about the idea of selective search. 
Tony made a blanket statement that the old idea of "selective 
pruning" (discarding "unpromising" moves from consideration) 
was dead, replaced by the new idea of "selective deepening" 3 or 
"selective extensions", namely, searching interesting moves more 
deeply. 

Could the idea of "selective deepening" be used to solve the 
problem that showed up in the Lachex-Bebe game? Well, it could 
if we had some way to detect that there was only one good move 
for the present position. I thought about it briefly, and reached the 
conclusion that there was no sure way to obtain the information 
without paying for it. In some very rare situations, the information 
is free, but we were talking about perhaps once in a million. I put 

3 Tony might have been the first person to use this phrase. 



54 Chapter 4 

the question in the back of my mind until the next day when I got 
on the plane back to Pittsburgh. 

The way people had been doing selective deepening or selec- 
tive extensions was mostly based on chess knowledge. If one side 
is in check, examine the position deeper. If a pawn is about to 
reach the eighth rank and become a queen, search the position 
deeper. In the case of ChipTest, it was very difficult to put in the 
chess knowledge for more selective extensions. But, if we could get 
the microprogram working, we would have the fastest chess pro- 
gram on the planet, and searching would be comparatively cheap. 
At least that was how the idea of "singular extensions" began. It 
was not really necessary to have a fast program. In any case, I rea- 
soned that we could afford to slow down the program a little bit by 
modifying the search or doing extra searches to gather information 
about whether a move is singular, or the only reasonable move. 
Once we verified that the move is singular, then we could search it 
more deeply. One way to gather the singularity information is to 
explicitly search all the alternative moves and to test whether they 
are significantly worse than the move in question. This could be 
done fairly efficiently along the "principal variation", or PV (the 
line where both sides play the best moves), as all the alternative 
moves along the PV have to be examined in a normal search any- 
way. This basic idea later became the "PV singular extensions". 
Anyway, I deemed the idea a theoretical curiosity at the time. I had 
more important things to attend to when I got back to Pittsburgh. 
It was about time to present my thesis proposal. 

Even though Ken Thompson had been on my thesis commit- 
tee for over a year, I had not yet presented my thesis proposal. After 
1 finished the design of the chess move generator chip, Professor 
Kung suggested that Dr Berliner should be on my committee. It 
seemed a good idea at the time that if I wanted to eventually build 
the ultimate chess machine, I would need the cooperation of at 
least one good chess player at some future date. Dr Berliner was 
without doubt the best chess player at Carnegie Mellon as well as 
a World Correspondence Chess (chess played by mail) Champion. 
His world title did not mean that he was of the top echelon in reg- 
ular chess, but he was still of International Master 4 strength, 
which would have placed him among the top fifty over-the-board 
players in USA. To be such a strong player is a considerable 
achievement in itself. He could play simultaneously against twen- 
ty players of my strength and win every single game. Ironically, 



The Chess Machine That Wasn't 55 

having Dr Berliner on the committee probably doomed any possi- 
bility of cooperation 5 . Anyway, I took Professor Kung's suggestion, 
and Dr Berliner became the third member of my thesis committee. 
Alan Fisher, a professor in the VLSI group, was the fourth member. 

When 1 got back from the ACM Tournament, Thomas had 
cleaned up and rewritten the search microprogram. It was working 
but not yet optimized, and there were still some hardware bugs 
that crashed the microprogram from time to time. I worked on 
both the hardware and the microprogram off and on while prepar- 
ing for my thesis proposal. A month after the ACM tournament, I 
presented my proposal, which was to complete the design of a real 
chess machine and work out the details of how to use multiple 
chess machines to get to unprecedented search speed. The thesis, 
when it was finished, became the blueprint of the future Deep Blue 
chess machine. 

Shortly after the proposal, I debugged the ChipTest hardware 
and optimized the microprogram. Thomas also had some free time 
to work on the software. Two months after the ACM Tournament, 
ChipTest played its first game against Hitech. The two teams were 
on cordial terms at the time and both were interested in getting a 
sparring partner 6 . ChipTest won the game, but Hitech was proba- 
bly still the stronger program overall. ChipTest, however, was 
clearly already one of the top chess machines in the world, just two 
months after its lackluster debut. It was searching about 300,000 
chess positions per second, approximately three times faster than 
any other chess program. Not bad for something that was designed 
as a chip tester. 

After ChipTest became a stable chess machine, I had some dis- 

4 From the top, the chess players can be ranked in the following order: World 
Champion, Grandmaster (GM), International Master (IM), senior master (SM), 
master, expert, class A, class B, class C and so on. The first three are internation- 
al titles. FIDE Master (FM) is a relatively new international title. An FM is roughly 
comparable to a senior master in playing strength. A Super Grandmaster is a very 
strong Grandmaster, but the term is mainly one of praise and not an official title. 

5 The arrangement created an unintentional public perception that the ChipTest 
team worked for Dr Berliner, which later became a source of friction. 

6 Dr Berliner had unrestricted access to ChipTest which he used as a sparring part- 
ner for Hitech extensively. We kept logs of all ChipTest games, so the sparring 
games played by Dr Berliner were of use to us as well. We also ran games against 
Hitech to debug ChipTest. Murray, who also worked on Hitech, conducted the 
debugging sessions for the ChipTest team. 



56 Chapter 4 

cussions with Murray Campbell about the idea of PV singular 
extensions. Not being a real chess player myself, I used Murray as 
a sounding board to find out whether the idea made sense. Murray 
was of the opinion that it deserved serious testing. Thomas imple- 
mented the idea during a weekend, and we pitted the version with 
PV singular extensions against a version without it in a twenty- 
game match. The result was a big surprise. The new version beat 
the old version by 15 to 5. A win-loss ratio of three to one corre- 
sponds to a difference of 200 rating points, or a full chess class. It 
is about the difference between the World Champion and a typical 
Grandmaster. It is huge. We were all very excited, although our 
excitement was tempered by the knowledge that the match result 
was between two closely related programs, and the strength differ- 
ence was probably greatly exaggerated. The twenty games were 
played from ten different starting positions with alternate colors. 
For the pair of games from the same starting position, usually the 
two programs would play exactly the same moves until a critical 
position where the new version managed to avoid an immediate 
disaster and ended up drawing or winning the game. Taking an 
analogy from 'Star Trek', the two versions of the program could be 
viewed as two identical Enterprise-class starships, one with all the 
long-range sensors disabled and one with all but the front long- 
range sensors disabled. As both starships head towards a massive 
black hole, the starship with the working front long-range sensor 
will be the one that detects the black hole, and escapes its deadly 
clutch in time. 

No self-respecting starship captain would allow his ship to be 
without the full complement of long-range sensors for long. What 
if there was an interesting planet starboard in the distance? 
Without the starboard long-range sensor, the planet would be 
missed completely. We found ourselves in a similar situation. PV 
singular extensions allowed the program to discover the potential 
disaster with the current best plan, but the extensions did not help 
the program find interesting alternatives to the current plan. In 
chess jargon, PV singular extensions does not help the program 
discover combinations, that is, non-obvious sequences of moves 
that lead to significant advantages. We needed a new idea. Could 
we extend the idea of PV singular extensions to attack the prob- 
lem? 

If an alternative chess move is not as good as our current best 
move, then our opponent must have refutations to every possible 



The Chess Machine That Wasn't 57 

continuation that we could play. To verify that a move was not the 
best move for us, we needed to find one working refutation move 
for our opponent against our move. What if our opponent had only 
one working refutation move after our move? Then it seemed to 
make sense to search deeper after our opponent's singular refuta- 
tion move. Intuitively, if a player has only one good move, it is 
possible that the move is the only one left that delays the 
inevitable disaster for the player. This is the main reasoning behind 
what we called fail-high singular extensions. (Computer chess 
researchers refer to refutation moves as fail-high moves.) 

The introduction of fail-high singular extensions produced a 
quantum leap in ChipTest's ability to detect combinations. But the 
newest version did not play measurably better than the older ver- 
sion in head-to-head competition against each other. The extra 
computation cost of the fail-high singular extensions wiped out 
the gain. We decided to keep the new extensions, figuring that 
they might have a positive effect against human players. At the 
very least, the extensions would make it impossible for the human 
players to guess whether the program would see a particularly deep 
combination. 

A couple of years later, Thomas decided to abandon his orig- 
inal thesis topic and do a detailed study of singular extensions and 
a few other search extensions that we used. There was some 
debate among the team whether some of Thomas' conclusions in 
his thesis should be taken at face value. Thomas' results were 
obtained using an unproven indirect measurement technique 
with Deep Thought, our next chess machine. His results were sug- 
gestive, but they provided more questions than answers. 
According to Thomas, singular extensions had little effect on pro- 
gram performance, at least as implemented in Deep Thought. It 
was not until 1991, when Deep Thought II became operational 
with an entirely new implementation of singular extensions, that 
we verified that the effects of singular extensions were quite sub- 
stantial. But this was after we introduced several additional new 
ideas, which were not present in Deep Thought, on how to con- 
trol the singular extensions and their derivatives. In particular, the 
Deep Thought II implementation also extended the search on 
moves that were not really singular. For instance, when the move 
was one of the only two good moves, Deep Thought II would still 
extend the search, although not as much as when the move was 
singular. 



58 Chapter 4 

Beating "World Champions" 

ChipTest was playing on roughly equal footing with Hitech after 
the introduction of singular extensions. It was tactically stronger 
than Hitech, but it had close to zero chess knowledge. We could 
have added some simple software evaluation to improve ChipTest's 
strength, but I had something else in mind. 

When I started the project, I planned to design an evaluation 
chip to go with the move generator chip. I had some rough idea on 
how the design would look, but I had no idea on what kind of 
chess evaluation function to put in. With the Hitech team around, 
I hoped that I could draw on the local expertise and experience. 
The problem was that Hitech was designed with a different philos- 
ophy, and was not compatible with my idea of fitting an entire 
evaluation function onto a single chip. By the time that ChipTest 
was up and running, I concluded that I needed to gain some design 
experience with chess evaluation function first hand before I could 
seriously entertain the idea of an evaluation chip. 

By then, the first field-programmable gate arrays (FPGAs) had 
just become commercially available. These FPGAs allowed logic cir- 
cuits of moderate complexity to be implemented with off-the-shelf 
components. They seemed to be ideally suited for the task of an 
experimental hardware evaluation function in a real chess 
machine. Professor Kung had some discretionary funds and agreed 
to foot the bill. This new machine would be mainly a learning 
experience on how to design chess evaluation hardware. But, based 
on the performance of ChipTest, the new machine could be 
expected to be a stronger chess player than any existing chess 
machine. I was also hoping to kick some butts at the 1987 ACM 
tournament! Given the crude nature of FPGA design software at 
the time, and the fact that I had never designed with FPGAs, the 
chance of entering the new machine in the 1987 tournament, 
however, was not good. 

One month before the 1987 ACM tournament, it became 
clear that there was absolutely no chance of getting the new 
machine ready in time. We drew up a contingency plan. First, we 
added to ChipTest the software evaluation of pawn structure that 
we had planned back in 1986. Second, we made sure that both the 
hardware and the software were in good shape by doing serious 
play testing, both against ChipTest itself and against Hitech 7 . 
Third, we "acquired" the fastest workstation that we could lay our 
hands on. We did not own the workstation — it was borrowed 



The Chess Machine That Wasn't 59 

through our connections with the Engineering Lab staff. Officially, 
the workstation was being tested before its installation in its real 
owner's lab. The faster workstation reduced the software overhead 
and ChipTest was able to search about 400,000 chess positions per 
second, a thirty percent increase in speed. 

After the addition of the software pawn structure evaluation, 
ChipTest became about 100 rating points stronger in games against 
machines. Since Murray was officially in the Hitech group, one big 
question was whether Hitech would be playing in the ACM 
Championship. It did not play in the 1986 tournament because its 
hardware was being modified. Dr Berliner decided to skip the com- 
petition again in 1987, citing the desire to concentrate on 
competitions against human players, so there was no problem 
with Murray continuing to work on ChipTest. 

Despite the absence of Hitech, the 1987 ACM tournament in 
Dallas had a strong field. The top finishers in the 1986 World 
Computer Chess Championship were all present with the excep- 
tion of Hitech. ChipTest was seeded third, behind Cray Blitz, the 
winner of the 1986 World Computer Chess Championship on 
tiebreak, and Belle, the winner of the 1986 ACM Championship. 
Bebe and Phoenix were seeded fourth and fifth respectively but 
tied for first at the World Computer Chess Championship. 

ChipTest beat Cyius 68K in round one easily. Round two 
brought a more difficult opponent, Lachex, which had finished in 
second place in the 1986 ACM tournament. ChipTest won a pawn 
early in the game, but it was a difficult endgame to win, going on 
until 3:00 am, Central Time, and was the last to finish. Pittsburgh 
is in Eastern Time so, biologically, it was 4:00 am for me and way 
past my bedtime. I was very tired but happy to have won the game. 

In the next round, we moved to the top board to play against 
Cray Blitz. We had Black. 

ChipTest played an opening called the center counter defense, 
in chess parlance. The defense did not have a good reputation. The 
main criticism of the center counter defense is the fact that the 
Black queen moves out into the open at a very early stage and can 
be subject to attacks. I grimaced when I saw the opening. One of 
the introductory chess books that I had read gave this particular 

7 We did not really do much play testing against Hitech ourselves. Dr Berliner was 
using ChipTest as a sparring partner against Hitech. We just checked out the 
game logs to make sure that nothing strange happened. 



60 Chapter 4 

opening as an example of how not to play in the opening phase of 
the game. Murray did not have a lot of time to prepare the open- 
ing, and decided to use the defense since it was relatively easy to 
prepare. Ken Thompson came by and asked, "How can you play 
this garbage again?" In the Lachex game, ChipTest had also played 
the center counter defense. Having seen the Lachex game, the Cray 
Blitz team prepared a better offensive line for White. 

Pretty soon after the opening phase, the board position 
looked a little bit dicey for ChipTest, and at this point it suddenly 
predicted that Cray Blitz would win a pawn. ChipTest went into 
what we called panic mode, trying to find a way out by going 
through all the moves. After about thirty minutes of calculation, it 
went for a sharp surprise continuation that would still lose the 
pawn but with some compensation. Mike Valvo, the tournament 
director, was quite shocked that any program would spend thirty 
minutes on a single move, and praised ChipTest for spending time 
wisely at a critical juncture. In reality, the reason why ChipTest 
spent thirty minutes on a single move was because singular exten- 
sions tend to cause the search tree to explode in size when used in 
conjunction with the Panic Mode. We had learned about this in 
our test games and therefore we gave ChipTest a very large time 
allocation whenever it was in panic mode. 

Given the surprise move from ChipTest, Cray Blitz inexplica- 
bly decided not to win the pawn. Had ChipTest played the normal 
continuation that Cray Blitz was predicting, Cray Blitz would 
have gone ahead and grabbed the pawn immediately. Cray Blitz 
had been calculating its response to the normal continuation dur- 
ing the entire 30 minutes of ChipTest's calculation. I breathed a 
sigh of relief. Cray Blitz's opening edge had dissipated. The game 
seemed to be about even for a few moves. But then, out of 
nowhere, ChipTest declared that it was winning. It had calculated 
a very long forcing line as a result of singular extensions and, was 
reporting that at the end of the line, it would win major material. 
Cray Blitz simply had no idea that its position was about to col- 
lapse like a house of cards. It took Cray Blitz three or four more 
moves before it realized that the game was lost. The game lasted 
only twenty-seven moves and was probably the shortest and 
worst loss of Cray Blitz's career up to that day. It was probably also 
the first public game where singular extensions played a major 
role. 

In the last round, we had White against Phoenix, a program 



The Chess Machine That Wasn't 61 

that ran on a network of workstations. We were thoroughly out- 
prepared in the opening. Phoenix had many more moves in its 
opening book than ChipTest. Since ChipTest had White, being 
out-prepared in the opening was not fatal. We did lose the advan- 
tage of the first move, and the game reached an endgame with 
equal possibilities for both sides. The author of Phoenix, Jonathan 
Schaeffer, was a master level chess player; after Dr Berliner, 
Jonathan might very well have been the strongest chess player 
among the active computer chess researchers. Former World 
Champion, Mikhail Botvinnik, was supposed to be working on a 
chess program, but he probably could not be considered an active 
computer chess researcher, as no one outside of his group had ever 
seen his program play a single move. Being a weak player myself, I 
listened intently to what Jonathan had to say about the position, 
and constantly checked his opinion against ChipTest's assessment 
of the position. Jonathan was later quoted as saying that I, being a 
weak player, was one of those guys whose emotions went up and 
down with their own program's evaluation of the position. This 
was certainly true in this game. Jonathan thought his program was 
in a better position, but with ChipTest thinking the position dead 
even, I was unconcerned. Jonathan was still in a cheerful mood 
when ChipTest's evaluation started to creep upward in its favor. 
Then it dawned upon Jonathan that Phoenix might be in trouble. 
Phoenix's position was no longer tenable. ChipTest had already 
seen ways to win material. A few moves later Jonathan resigned for 
Phoenix, and ChipTest won the Championship with a clean 
sweep. 

ChipTest was probably at this point the best chess machine in 
the world. It had won the ACM title ahead of three of the top four 
finishers in the 1986 World Computer Chess Championship. 
Hitech was the only one among the top four at the World 
Championship that was missing, and ChipTest had held its own 
against Hitech before the ACM tournament. How strong was 
ChipTest? I have reasons to believe that it might have been the first 
senior master machine. ChipTest was never fully rated in human 
tournaments, so this is just a guess. 

ChipTest received a $2000 prize for the win, which was more 
than the estimated $500 cost of building it. To celebrate the win, 
we used part of the prize money to hold a TG party for the whole 
Department in ChipTest's honor. It seemed very appropriate, given 
how ChipTest was born in the first place. 



62 Chapter 4 

An Invitation from California 

During one of the conversations at the 1987 ACM Championship, 

Jonathan Schaeffer, the author of Phoenix, said that there were 

two people whom he really looked up to in the computer chess 

field. 

The first person was Ken Thompson. Jonathan respected Ken 
not just for his achievement and contribution to computer chess, 
but also for his unselfish acts. He mentioned a few of Ken's deeds, 
one of which I observed myself. Each year at the ACM 
Championship, Ken would collect the game scores after each 
round, enter them into the computer, and then post them to the 
rest of the world. Not something you would expect from a busy 
and important person who had won the Turing Award, the equiv- 
alent of the Nobel Prize for a computer scientist. 

The second person that Jonathan looked up to came as a big 
surprise to me. It was Dr Hans Berliner. Looking back, I was sur- 
prised partly because of my professional differences with Dr 
Berliner, and partly because I considered him not the easiest per- 
son to work with, but I guess I can understand why Jonathan 
thought highly of him. When Jonathan looked at Dr Berliner, he 
must have seen quite a bit of himself, or at least something of what 
he would like to be. They were both very good chess players doing 
research in computer chess. Jonathan was still an up-and-comer, 
while Dr Berliner was well established. Dr Berliner had worked on 
computer chess since 1970 when he entered his first program in 
the ACM competition. Having garnered the title of World 
Correspondence Chess Champion earlier, Dr Berliner may have 
hoped that he could be the person who would finally solve the 
Computer Chess Problem, creating a chess machine that could 
beat the World Chess Champion in a match. He went back to grad- 
uate school at a fairly advanced age, earned a PhD degree, and 
became a research faculty member at Carnegie Mellon University, 
where the Computer Science Department had a history of pio- 
neering work in computer chess. He gathered together a group of 
graduate students with varying skills and led them in the creation 
of Hitech. There are very few people who could match Dr Berliner's 
drive. 

A few weeks after my conversation with Jonathan, the friend- 
ly competition between the ChipTest team and Dr Berliner 
suddenly turned into an open rivalry. 

Before I tell you the story, I need to put you in the right frame 



The Chess Machine That Wasn't 63 

of mind. First, the story is important. Things like this do happen. 
I don't enjoy telling it, but the events in the story marked an irre- 
versible transition in the relationship between the ChipTest team 
and Dr Berliner. Second, the story may make you look at Dr 
Berliner unfavorably. Before you do that, please think twice about 
what you would have done in his place. Imagine that you have a 
lifetime goal, a higher calling. You have worked for decades to 
reach the goal, making many personal sacrifices. You are now a 
widely recognized authority in your chosen field. You have 
achieved great successes in the last few years and your goal seems 
to be in sight. Now imagine a bunch of students coming out of 
nowhere. They are not from your field and they have different 
ideas. You have little persuasive power over them since they are 
not your students. Their approach is the exact opposite of what 
you are advocating. To make matters worse, they seem to be mak- 
ing rapid progress, which makes it even harder to persuade them 
to follow your vision. What would you do? What can you do? It is 
a very tough situation. 



Andreas Nowatzyk, my office mate, was spending time at NASA's 
Jet Propulsion Lab QPL) in Pasadena, California. He met Stuart 
Cracraft, a computer chess enthusiast, who happened to be work- 
ing at JPL at the time. Stuart asked Andreas a lot of questions about 
ChipTest. A big chess tournament, the American Open, was about 
to take place in Los Angeles. Stuart apparently had good contacts 
with the tournament organizers and suggested that we enter 
ChipTest into the tournament. Stuart mentioned that he had 
friends who would gladly help with setting up communications 
and operating the machine remotely. Andreas thought that it was 
a good opportunity to see how well ChipTest would do against 
human players. 

Thomas was all for entering the tournament. Murray did not 
want to take sides, as both he and I knew that Hitech was going to 
the same tournament. I was against participating, but not because 
Hitech was playing. I found that I was spending too much time 
watching ChipTest play and not getting much work done. I was 
not quite as disciplined as I would have liked. If ChipTest started 
playing in tournaments, the new machine that I was working on 
might never see the light of the day. Anyway, I sent an e-mail to Dr 
Berliner asking for his opinion. 



64 Chapter 4 

The next morning, Dr Berliner sent back a message vehe- 
mently against ChipTest going, giving three reasons. The first was 
that ChipTest was totally untested against human players. The sec- 
ond was that Carnegie Mellon's reputation could be hurt. His third 
was that it was foolish to let an outsider, not affiliated with the 
university, operate ChipTest. 

From my point of view, ChipTest being untested against 
human players would have been the perfect reason to take it out 
for a spin. We were in a research university after all. Furthermore, 
I had never heard of a university's reputation ever being ruined by 
student projects doing poorly in chess tournaments. Dr Berliner's 
third objection brought up a valid concern, but ChipTest was a stu- 
dent project with a zero budget, and 1 did not see any strong 
reason against getting outside help to operate the machine. 
Otherwise, we would have to shell out a serious amount of our 
own money to fly to Los Angeles 8 . 

I disagreed with Dr Berliner's reasoning, but since I did not 
really want to enter ChipTest into the tournament I just forwarded 
Dr Berliner's message to Andreas. 

Andreas replied in the afternoon that the potential operator 
did have an affiliation with CMU. He was Jim Gillogly, one of the 
early pioneers in computer chess and a PhD from the Department 
at roughly the same time as Berliner. A couple of hours later, Jim 
sent me an e-mail formally agreeing to be the operator, if needed. 
I passed this information on to Dr Berliner. 

Dr Berliner was probably taken off guard. This time he came 
up with a different reason. The ChipTest team had not published 
the paper on singular extensions yet, but Dr Berliner knew about 
them. The Hitech group was just in the process of getting singular 
extensions to work on their machine. Since Hitech did not have 
Singular Extensions, Dr Berliner worried that it could finish behind 
ChipTest, and the Hitech group would look bad. Well, it was not 
completely unreasonable. We had come up with the idea of singu- 
lar extensions ourselves without any help from Dr Berliner, so in 
that sense his request was unreasonable. On the other hand, we 
owed Dr Berliner something. When Murray agreed to help Thomas 
and me with the evaluation function for ChipTest, he did not start 
from scratch. Unbeknown to me at the time, Murray used the 

8 The Department did pay for the trips to the ACM Championship, but compet- 
ing in the ACM Championship was an academic activity. 



The Chess Machine That Wasn't 65 

"Cray Blitz Simulator" code 9 written by Dr Berliner as the starting 
template to create ChipTest's evaluation function. The Cray Blitz 
Simulator was a simple (compared to Hitech's) evaluation function 
used by Berliner to run on Hitech to simulate the expected behav- 
ior of Cray Blitz. The amount of code from Dr Berliner was less 
than a half of one percent of ChipTest's total code, but it was his 
nonetheless. I first learned about the nature of the code in ques- 
tion only a few weeks before the 1987 ACM Championship. 
Gordon Goetsch, who was in the Hitech group, was joking about 
the possibility of Cray Blitz being outplayed by the Cray Blitz 
Simulator. 

If Dr Berliner had stated this objection right at the beginning, 
we probably would have just honored his request. By this time, I 
was a little bit mad at him. I had a discussion with Thomas and we 
came up with a compromise. We would go to the tournament only 
if Hitech has the singular extensions working. In reality, we were 
not making any travelling plans as we could not imagine they 
could have them working in time. 1 informed both Berliner and 
Andreas of the decision. 

Andreas surmised that another possible reason for Dr 
Berliner's objection, may have been because Hitech was within 
sight of a major prize and Berliner did not want to be beaten at the 
post — a perfectly understandable stance given that the two teams 
were potential rivals. Back in the late 1970s, Professor Edward 
Fredkin set up a $100,000 prize for the first computer to defeat the 
World Chess Champion in a match. Carnegie Mellon University 
was entrusted with managing the Fredkin Prize, and Dr Berliner 
was the chairman of the Fredkin Prize Committee. To encourage 
steady progress in the field, two additional but smaller prizes, the 
Fredkin Intermediate Prizes, were set up. The Belle team had 
claimed the first Fredkin Intermediate Prize for master level per- 
formance. Perhaps Dr Berliner had set his eyes on the second 
Fredkin Intermediate Prize for Grandmaster level performance ever 
since Hitech had been built. Not fully aware of the rules for the 



9 How did ChipTest end up using this code from Dr Berliner? None of us knows 
for sure any more. Thomas Anantharaman believes that he might have retrieved 
the code after Gordon Goetsch suggested it to him. Gordon, in turn, might have 
been simply trying to help us make it to the 1986 ACM Champion in the short 
seven weeks that we had. Gordon, however, could not confirm nor deny what 
Thomas said. 



66 Chapter 4 

award, Andreas pointed out that, with singular extensions, Hitech 
might just be good enough to achieve a 2400 rating and become 
the first computer senior master. However, the rules for the second 
Intermediate Prize specified a twenty-five-game performance of 
over 2500. I did not believe Hitech would be good enough, even 
with the singular extensions. Neither did I believe that ChipTest 
could do it. If Dr Berliner wanted Hitech to be the first official sen- 
ior master then good luck to him. I had bigger fish to fry. Anyway, 
we had agreed that we would not go, although with a proviso. 

For Dr Berliner, Hitech happened to be his last chess 
machine — it took years to complete and stayed in service for over 
a decade afterwards. ChipTest took seven weeks to put together, 
and was only one year old at the time. But ChipTest was effective- 
ly fully depreciated, having earned more in prize money than it 
had cost, and I had a new machine in the pipeline. ChipTest miss- 
ing a tournament was not as important to me as getting the new 
machine finished. 

Shortly afterwards, Professor Kung called for a meeting to dis- 
cuss things and both Dr Berliner and I attended. Murray and 
Thomas were not present. The first part of the meeting was to talk 
about the invitation from California. Dr Berliner made his state- 
ment first, expressing his dissatisfaction that ChipTest was using 
the "Cray Blitz Simulator" code which was a surprise to me. I had 
never seen the code until after the 1986 ACM Championship but 
once I knew its origin I thought that Murray used it with Dr 
Berliner's blessing. For most of the year before, I had assumed that 
Murray wrote the code himself. Had Dr Berliner told us about his 
displeasure, we would have gutted the code and rewritten things 
from scratch. As it was previously noted, it was less than a half of 
one percent of the total code anyway. I explained that I knew 
nothing about it, and Murray was the one who knew anything 
about the code. (I did not know it at the time, but Thomas might 
have been the one who retrieved the code.) Dr Berliner accepted 
my explanation. I then reiterated that we did not want to play in 
the tournament, so the problem with the invitation was finally 
over. 

The second part of the meeting was about my future. The 
Department had been mulling over the possibility of retaining me 
after my graduation. Naturally, they would prefer that Dr Berliner 
and I work together. I did enjoy life a great deal in the Department, 
but I had one major concern. What about the credit? Dr Berliner 



The Chess Machine That Wasn't 67 

answered "The credit will take care of itself". I was not sure exact- 
ly what he meant by this, but I had seen what happened with 
Hitech. Most people referred to it as Dr Berliner's machine and Carl 
Ebeling was largely unknown to the outside world. This perception 
may well have been due to inaccurate reporting in the press, but 
the assumption stuck and I would not want anything similar to 
happen on any project I worked on. Well, it was an interesting 
offer, but I had to decline, since I had no confidence that the cred- 
it would indeed take care of itself. After graduation I would have to 
look for a job elsewhere to continue the work. The job prospects 
for a computer science PhD were not good at the time. Oh, well, 
what will be will be. The thing to do would be to make the new 
machine as successful as we could, in particular, to go after the 
Fredkin Intermediate Prize. We also needed to make it obvious to 
anyone that the new machine was completely separate from the 
work of the Hitech team. 

I became a lot more sensitive to the credit issue after this 
episode. I didn't want to be mistreated, and I didn't want to see 
people mistreated. I swore after the meeting that our team would 
not have an official leader. The team was functionally partitioned, 
and it would be unfair to any one to declare a particular person as 
the leader. 



Looking back, I wonder whether I was just lucky. All my project 
experience up to that point had been some sort of partnership — 
my undergradate project in Taiwan and the GE project had both 
been with partners. The attitude of my faculty bosses had been 
hands off and just letting us do whatever we pleased. So the deci- 
sion to have no official leader was kind of natural. The chess 
project turned out to be a project that required complete dedica- 
tion from every member of the team. This dedication was possible 
only because we were equal partners. 

Several books and articles stated that the Hitech team and the 
ChipTest/Deep Thought team were bitter enemies. The relation- 
ship was more civil than that. For the Deep Thought team, it was 
easy — we ended up on the winning side. Also, things only really 
turned bad after the invitation from California. It is safe to say that 
neither team would invite the other to their private parties after- 
wards. After we left Carnegie Mellon, time and distance helped to 
heal the wounds; I think both sides mellowed. When Deep 



68 Chapter 4 

Thought II became active, we no longer considered Hitech as a 
competitor. We still did not like the idea of losing to Hitech. Hell, 
we did not like to lose to anybody, Garry Kasparov included! 



A 



CHAPTER 5 



The Race for First Machine Grandmaster 



Some 2200s 

During the 1970s and 1980s, the British science fiction writer 
Douglas Adams wrote a highly popular science fiction trilogy that 
the BBC turned into both a radio show and a television series. 
While I was serving in the Taiwanese Army, a local radio station 
was carrying the BBC radio show based on the first book, The 
Hitchhiker's Guide to the Galaxy. However, I did not get to read the 
books themselves until I lived in the United States. 

In the universe of the Hitchhiker's Guide, an alien race built 
a computer called Deep Thought, to solve the question of "Life, the 
Universe, and Everything". After eons of computation, it came up 
with the answer "42". But no entity, including Deep Thought, 
knew the original question any more. Deep Thought, however, 
being the most powerful computer ever built, knew how to build 
an even more powerful computer which would answer the ques- 
tion, of what was the question, to the answer "42". 

Well, a machine that could defeat the World Champion 
would surely be worthy of the name Deep Thought. 

I did not believe that the machine we were building in early 
1988, which was also our first real chess machine, would be suffi- 
cient to defeat the World Champion. It was probably at least 100 
times too slow. Factoring in the speed discrepancy, the first version 
of the new program was called Deep Thought 0.01. It later became 
Deep Thought 0.02 after we got a dual-processor version working. 
Eventually, when we decided to join IBM and build a new machine 
with a new name, the name became simply Deep Thought. 

Andreas came back to Carnegie Mellon from Pasadena after 
the New Year. We talked off and on about the possibility of creat- 

69 



70 Chapters 

ing a program that could tune the chess evaluation function auto- 
matically. Andreas had an exciting new idea. The new idea, 
however, required that the evaluation function be written in a dif- 
ferent way. Murray liked Andreas' idea and agreed to write a 
completely new evaluation function in the way Andreas specified. 
The new evaluation function was written in part to handle the new 
evaluation primitives supported by the Deep Thought hardware, 
which were about to be completed. The original reason for the new 
evaluation function, and the automatic tuning software, however, 
was to start from a clean slate. We needed to avoid any of the prob- 
lems with the "Cray Blitz Simulator" code as we had had with 
ChipTest, so that we would be able to compete freely. The auto- 
matic tuning of the evaluation function was often mentioned as 
the most novel aspect of Deep Thought, but it came purely out of 
necessity. Andreas helped quite a bit in bringing up ChipTest earli- 
er, and with this increased involvement I finally got him to agree 
to be seen as one of the authors of Deep Thought. 

Once Andreas became seriously involved, I set my sight on 
Mike Browne, my other office mate. (Okay, being my friend has its 
drawbacks.) One of the biggest problems for ChipTest was its open- 
ing book, or rather, its lack of opening book. Mike agreed to write 
a program that automatically examined all the opening lines to 
look for opening traps (bad opening moves), and theoretical nov- 
elties (new opening moves that are interesting), and became the 
fifth member of the Deep Thought team. 

During most of the 1980s, the Fredkin Foundation sponsored 
a series of annual computer chess events. The final event of the 
series took place in 1988, the year after ChipTest's win at the ACM 
championship. The money used to sponsor the events came from 
the interest accrued from the prize fund for the $100,000 Fredkin 
Prize. The Fredkin events, unlike the computer competition in the 
ACM championship, pitted the top chess programs against human 
chess players. Traditionally, the reigning ACM Computer Chess 
Champion was invited along with possibly some of the stronger 
programs at the time. The 1988 final edition was the Fredkin 
Masters Open, from May 28 to 30 on the Carnegie Mellon campus. 
It would be a six-round tournament, with two rounds per day. 
ChipTest, as the reigning ACM Champion, was invited, along with 
Deep Thought 0.01. Hitech, as the 1985 ACM Champion, was also 
invited. Hitech would be using Singular Extensions for the first 
time in tournament play. 



The Race for First Machine Grandmaster 71 




The two versions of Deep Thought 1 . 

The wire-wrapped (top) and the printed circuit (bottom) versions of the Deep 
Thought 7 board. The two wire loops on the right side of the wire-wrapped 

board are used to delay two critical signals by a few nanoseconds. (Electrical 
signals tiavei slightly less than one foot over the wires eveiy nanosecond.) 



72 Chapter 5 

When the work began on the Deep Thought hardware, I was 
hoping to get it ready for the 1987 ACM Championship but Deep 
Thought never made it. I was not going to miss the Fredkin 
Masters Open; however, I ended up cutting it very close. The most 
important new feature in Deep Thought, compared to ChipTest, 
was its hardware evaluation function which recognized many 
dynamic positional features, including chess concepts such as 
pawn structure, passed pawns, rooks on open files, and so on. The 
new hardware evaluation was implemented with field programma- 
ble gate arrays that took longer to program than I expected. The 
wire-wrapped Deep Thought board also had some nasty electrical 
problems that took a while to track down. Furthermore, the Deep 
Thought board, besides having the new hardware evaluation func- 
tion, had two chess processors, and both needed to be fully tested. 
With the additional complexity, the cost for the Deep Thought 
board was close to $5000, or ten times the cost of ChipTest. Just 
procuring the parts for Deep Thought took longer than the time to 
design and build ChipTest. By March 1988, the Deep Thought 
board was partially working, but the debugging of the board, along 
with the creation of the new software, seemed to take forever. Two 
days before the Fredkin Masters Open, the Deep Thought board 
was still being rewired, and Thomas continued to add finishing 
touches to the new software until the event started. 

The Pittsburgh Chess Club ran the tournament. About thirty 
players participated, approximately twenty at the master level or 
above. One of the out-of-town masters was Alexander Ivanov, a 
recent Soviet emigre. Alexander was rated 2597 on the US rating 
scale, which was close to Grandmaster strength, even though he 
did not have a title at the time. (He became a Grandmaster about 
a year later.) The top Pittsburgh master was Vivek Rao, a teenager 
rated at 2491 who also happened to be the top player in 
Pennsylvania. Vivek's rating put him at close to International 
Master (IM) strength. The masters from Pittsburgh were thought to 
be a particularly tough crowd for computers. Many of them had 
participated in earlier editions of the Fredkin events or played 
Hitech before. It would be a very good test for the computers. 

I paid the tournament entry fees and US Chess Federation 
(USCF) membership dues for both ChipTest and Deep Thought, 
using ChipTest's prize money from the ACM Championship. After 
paying, I checked the wall chart for the participants. Neither 
ChipTest nor Deep Thought was rated at the time, and I was sur- 



The Race for First Machine Grandmaster 73 

prised to see ratings placed next to the their names. I asked the 
tournament director, and he said that the ratings for both 
machines were Dr Berliner's estimates. The programs were both 
listed as 2200, essentially very weak masters. Hitech was rated 2376 
at the time. I winced but did not say anything as I could under- 
stand the reasoning for these ratings. The low estimate ratings in 
the strong field meant that for both ChipTest and Deep Thought 
our first round opponents would be strong masters. 

Thomas was the operator for ChipTest and I operated Deep 
Thought. Deep Thought's first opponent Ross Sprague, rated 2339 
on the wall chart, was a lawyer from out of town who played con- 
servatively. The game was still undecided when it was time to 
adjourn till the next morning. Meanwhile, ChipTest scored its first 
win against a human opponent, a strong master. 

I came back early the next day to finish the game with Ross. We 
both knew that the game was heading to a draw but that Ross could 
try to avoid the draw by playing a risky move. But first, he played on 
for a while to see whether Deep Thought would self-destruct. Finally, 
he figured that a 2339-rated human player should beat a 2200 
machine and decided to take the risk. Deep Thought's evaluation 
immediately jumped up, and a few moves later, Ross threw in the 
towel. Deep Thought probably would have drawn the game, if not 
for the 2200 rating estimate. In the second round games, both Deep 
Thought and ChipTest drew against strong masters. 

Deep Thought's third round opponent was Kimball Nedved. 
Kimball's son, Rudy, worked in our department, and Kimball him- 
self was quite familiar with chess computers. He had both beaten 
and lost to Hitech before. Kimball was sufficiently comfortable 
that he would make comments from time to time during the game. 
He had drawn ChipTest in the previous round and was compli- 
mentary of ChipTest's play. Deep Thought was not quite stable yet, 
and all of a sudden, it played an instant move due to an internal 
inconsistency. I raised my eyebrows when I played the move. 
Kimball had no idea why I did this, and said questioningly, "Why? 
That is a good move." Well, it just happened to be. There were no 
more incidents in the game, and Kimball resigned without ever 
realizing that Deep Thought could run into the same problem 
again. Meanwhile, ChipTest won its third round in a hair-raising 
game. The fourth round saw Deep Thought winning another game 
and ChipTest drawing its game. We would be meeting the top play- 
ers on the third day. 



74 Chapter 5 

In the fifth round, Deep Thought was paired against 
Alexander Ivanov, and ChipTest against Vivek Rao. Deep Thought 
was still using the same opening book that ChipTest used in the 
1987 ACM Championship, and that, of course, meant we were see- 
ing the center counter defense, also known as the Scandinavian 
defense. I was surprised to see Alexander, the soon-to-be 
Grandmaster, spending a lot of time in the opening. He did say 
before the game that he had never studied the Scandinavian 
defense seriously. Maybe the center counter defense was not 
garbage after all. Alexander played carefully and powerfully. Then 
he went into deep thought over a position that looked quite dan- 
gerous for Deep Thought. It appeared to me that Alexander could 
unleash a very strong attack on Deep Thought's king. To my sur- 
prise, Alexander played a move that seemed to release all the 
tension in the position. What was he doing? It looked like an even 
ending to me. On a closer look at the board, I began to get a sink- 
ing feeling. It wasn't an even ending. Alexander had seen deeper 
than Deep Thought, and Deep Thought lost its first ever game. 
While Deep Thought was having its problems, interesting things 
were unfolding on ChipTest's board. Before Vivek's game with 
ChipTest, he was openly expressing his contempt of chess-playing 
computers, having had numerous pleasant memories at the 
expense of Hitech from earlier encounters. ChipTest forced Vivek 
to resign in twenty-two moves with an unexpected sacrifice. Hours 
after the game was over, Vivek was still showing other people 
ChipTest's surprise combination. 

In the last round, ChipTest's undefeated record against 
human chess players came to a screeching halt, losing to a strong 
master. Despite Vivek's loss to ChipTest, he continued to express 
his contempt of chess computers, and got a chance to prove his 
point in the last round. He was paired against Deep Thought. 
Now chastened by ChipTest's tactical ability, Vivek went for a 
quieter game against Deep Thought. An "isolated queen pawn" 
position was reached. Deep Thought had more control in the 
center of the board, but also a potential weakness in its isolated 
queen pawn. Deep Thought's queen pawn did not have any 
friendly pawns on the adjacent files. If Vivek could drive the 
game into an ending with most of the pieces traded off, he would 
very likely win, as Deep Thought's queen pawn would become a 
major endgame weakness. Vivek managed to trade off two pieces, 
and seemed to have obtained a favorable position. Then Deep 



The Race for First Machine Grandmaster 75 

Thought started playing dumb looking moves. A smile formed on 
Vivek's face. Did he sense a kill? Deep Thought remained uncon- 
cerned. Vivek continued his quiet maneuver with innocent 
looking moves. Then Deep Thought pushed its pawn on f file, 
threatening to open up the position. Its score jumped. Vivek 
froze up, took another look, and left the room for fresh air. It was 
a good move. He came back and sat at the table, this time with 
his hands over his ears. Someone opened the door and made a 
slight noise. Vivek hushed the guy. After thinking for twenty-five 
minutes, Vivek made his move. A series of piece exchanges 
ensued. After all the smoke cleared ten moves later, Vivek was 
down a pawn. A few moves later, when he was about to lose more 
material, he resigned. So the top player in Pennsylvania lost 
twice in the same day to the newest USCF players in 
Pennsylvania. 

Ross, our first round opponent, dropped by after the Vivek 
game. He commented, "Some 2200s." He then added that if he had 
known what he was dealing with, he would have just taken the 
draw. 

Alexander Ivanov won the event, scoring 5 points out of 6. 
Deep Thought tied for the second place with two masters, scoring 
4.5 out of 6. ChipTest tied for fifth place, scoring 4 out of 6. Hitech 
had a bad tournament, scoring 3.5 out of 6. Deep Thought's per- 
formance rating was 2599 for the tournament, while ChipTest's 
was at 2521. Hitech performed at 2312. 

If ChipTest and Deep Thought could be considered as a single 
machine, then its twelve-game performance would be around 
2560. The rules for the Fredkin Intermediate Prize called for a 
twenty-five-game performance of over 2500. We had a very good 
shot, assuming that Hitech would not do it first. Before the Fredkin 
Masters Open, I did not believe that either Hitech or ChipTest 
could have done it. But it now looked like ChipTest could do it, 
and who could say for sure that Hitech could not do it too? We had 
a new sense of urgency. Deep Thought had the best shot of the 
three machines, but it would be a race. We had the lead and the 
stronger machines. But don't forget what happened to the hare in 
the race against the tortoise. 

Thank youl Thank you! 

One of the drawbacks of constantly bringing new machines to 
chess competitions, is the need to deal with the numerous exciting 



76 Chapters 

bugs waiting to happen. ChipTest had a memorable first outing at 
the 1986 ACM Championship. The debut of Deep Thought 0.01 in 
the Fredkin Masters Open certainly had its tense moments, but 
there was nothing obvious, at least from our opponents' view- 
points. Deep Thought 0.02 played its first tournament at the 1988 
US Open held in Boston from August 7 to 19. It was an experience 
reminiscent of ChipTest's first tournament. 

Deep Thought's wire-wrapped card had two chess processors 
on it. At the Fredkin Masters Open, only one of the processors was 
in use; there was no time to get a dual-processor version working. 
The single-processor version that played at the Fredkin Masters 
Open was adapted from ChipTest, using the old search code but 
with the brand new evaluation function. For reasons now lost in 
antiquity, we did not bother to play in tournaments in the two 
months leading up to the US Open. Perhaps, we wanted to get the 
dual-processor version as soon as possible. The other possibility 
was that we wanted to get a decent opening book. 

The ChipTest opening book was grossly inadequate for play- 
ing against human players. Murray Campbell received his PhD 
degree before the Fredkin Masters Open, and was offered the fac- 
ulty position of Research Associate at the CS Department. The 
position was equivalent to Assistant Professor but in a research 
track. Berliner was his semi-official boss '. After the Fredkin Masters 
Open, Murray was told that he was not to spend any time on either 
Deep Thought or ChipTest — after all, he was employed to work on 
other projects. Murray followed the directive to the letter in office 
hours, but continued to work on the two programs in his leisure 
time. With Murray's time becoming a scarce resource, we moved 
the new opening book to the top of the to-do list. Since Murray 
would not have enough time to work on it, we sought help from 
outside the university. International Master, Larry Kaufman, 
offered to provide us with the opening book that he had prepared 
for the Rex chess program, and we took him up on his offer. 
Ironically, Rex had been ChipTest's first victim at the 1986 ACM 
Championship. 

While we were bringing up the new version, Hitech played in 
the 1988 Pennsylvania State Championship from July 23-24, won 
the title, finally getting its rating over 2400, and became the first 

1 Dr Berliner was a Senior Research Scientist, which was equivalent to a full pro- 
fessor but in a research track. 



The Race for First Machine Grandmaster 77 

official computer senior master, three years after its birth 2 . Both 
ChipTest and Deep Thought were higher rated than Hitech, but 
had not yet played twenty rated games, the minimum needed to 
be official senior masters. While there was no prize associated with 
the senior master title, there were some misgivings among the 
Deep Thought team members, that we should have sent out the 
single-processor version of Deep Thought to tournaments. 
Anyhow, Hitech the tortoise was catching up, and we could not 
wait any longer. 

The US Open, the showcase event for the US Chess 
Federation, was a twelve-round tournament held in Boston. 
Thomas Anantharaman had a friend with whom he could stay in 
Boston, so he volunteered to drive to Boston, and be the operator. 
Thomas' expenses would be paid from our earlier winnings in the 
ACM Championship and the Fredkin Masters Open. As computers 
were not allowed to get the prize money in regular chess tourna- 
ments, we would not be getting any more winnings, and had to 
watch how much we spent. We would be playing with the dual- 
processor version for the first time. 

The first three games at the US Open were played with the 
dual-processor code, and the results were disastrous. The first game 
was drawn as a result of a bug at the interface between the hard- 
ware and the software code, although we were not certain of the 
cause at the time. Deep Thought 0.02 managed to win the second 
game when the opponent resigned early, but had he known what 
was to happen in the third round, he probably would have played 
on. In the third round, we got an expert-level opponent who sim- 
ply refused to resign when his position was clearly lost. Deep 
Thought 0.02 was up by a rook, but the opponent played on. Then 
strange things started to happen. Deep Thought 0.02 started play- 
ing nonsensical moves. The opponent made a move threatening to 
checkmate it on the next move, and Deep Thought 0.02 did not 
bother to defend against the threat. The opponent just made the 
mating move, and then exclaimed "Thank you! Thank you!" as he 
ran off. 

It turned out that Thomas' new dual-processor code was 
accessing the hardware in a different way from the single-proces- 
sor code. This had the side effect of causing the hardware 

2 In the first edition of his book Kasparov vs. Deep Blue, Monty Newborn erro- 
neously stated that Hitech became a senior master in 1987. 



78 Chapters 

microprogram to reverse the sign of the value for checkmates 
under certain conditions. Instead of avoiding being checkmated, 
Deep Thought was enjoying being checkmated! 

With Thomas in Boston, it was difficult to debug the dual- 
processor code, and we reverted to the more reliable 
single-processor program after round three. In the next five games, 
Deep Thought 0.01 allowed one draw, winning the other four. 

Round nine brought the first major opponent, International 
Master Igor Ivanov. (Years later, I picked up a book on the early 
career of World Champion Garry Kasparov, and was surprised to 
read about a tournament where Igor and Garry tied for first place. 
But I am getting ahead of myself.) I had seen Igor in action much 
earlier, even before ChipTest's existence. It was a tournament game 
between him and Hitech, played on the University of Pittsburgh 
campus, a ten-minute walk from my office. In that game, Igor 
effortlessly arranged his pieces for an attack and slaughtered 
Hitech. Gordon Goetsch of the Hitech group pointed to this game 
as an example of how far computers had yet to go. After ChipTest 
was built, I tested out positions from the Igor vs. Hitech game and 
was quite impressed with his play. ChipTest probably saw Hitech's 
demise a few moves earlier than Hitech did, but also had no idea 
what he was doing. Igor was of Grandmaster strength, but he did 
not have a Grandmaster title. He had defected from the USSR dur- 
ing a tournament held in the West, and this might have made it 
difficult for him to obtain the title. He was the Canadian co-cham- 
pion, and at 2641 on the US scale was one of the top players at the 
Open. 

How would Deep Thought do against Igor Ivanov? In 
Pittsburgh we were obviously nervous. In Boston, Thomas was 
probably ignorant of his history but his rating alone would be suf- 
ficient to impress. Igor had Black. Later, Thomas mentioned that 
Igor was reading a magazine during the early part of the game; he 
had no respect whatsoever for Deep Thought. The results from the 
Fredkin Masters Open were still largely unknown in the chess 
world. 

The game against Igor turned into one of the strangest affairs 
that I have ever seen. Deep Thought liked to push its pawns. In 
chess terms, it liked to increase space. Good chess players are 
taught at an early stage that pushing pawns too far could create 
weaknesses. The "hypermodern" chess school of thought is to 
deliberately provoke the opponent to push their pawns, hoping 



The Race for First Machine Grandmaster 79 

that the pawns become over-extended and difficult to maintain. If 
that was Igor's intent, he succeeded brilliantly, at least for the pro- 
voking part. By move nineteen, all of Deep Thought's pawns had 
advanced at least one square; in fact, with the exception of the two 
pawns on the edges, all of them had advanced by at least two 
squares. With Deep Thought pushing pawns with abandon, I was 
quite concerned, but it really liked its position. What was going 
on? According to Thomas, by this time, Igor had put down his 
magazine for several moves; he had seen something that he did 
not like. Deep Thought's next move pushed all of his pieces back 
to the first two ranks, and he could hardly move. Its pawns were 
not over-extended — they were dominating the position. Two 
moves later, Igor's position deteriorated and became critical. On 
move twenty-nine, he resigned just as he was about to suffer heavy 
material loss. 

The 1987 US Open Champion, Grandmaster Lev Alburt, rated 
slightly higher at 2652 than Igor, was paired against Deep Thought 
in the tenth round. Igor's loss to Deep Thought had been a big 
upset, and Lev was adequately warned about what he would be fac- 
ing. Lev played cautiously and entered the endgame with a 
promising position. On move nineteen, Deep Thought's position 
became critical, and it pushed a pawn in front of its king by two 
squares. Lev was surprised by the move which was probably the 
only move that held the position for Deep Thought. While Deep 
Thought's position looked ugly, it was not losing. The wily Lev, 
however, found a way to make progress. The game then became a 
rook ending where both sides had only one rook and five pawns 
each. Lev still held the edge, but rook endings are notoriously dif- 
ficult to win. Perhaps we could get a draw. Deep Thought indeed 
had a repetition draw in hand by move fifty-two, if only it had 
played the drawing move. For reasons unknown, Deep Thought 
found the move, but then decided to look at an alternative draw- 
ing move. The new move, however, lost outright. Deep Thought 
found that the new drawing move did not work, but the software 
controlling the time allocation was not smart enough to allow it to 
switch back to the real drawing move. All of us watched in agony 
while Deep Thought spent twenty minutes trying to find the draw- 
ing move that it already knew about. It did not rediscover the 
move, and we resigned for the computer on move sixty-one. 

Deep Thought won the next game against a master player, 
and then drew the final round game against a senior master. Its 



80 Chapter 5 

provisional rating dropped from 2599 to 2495, mostly as a result of 
its buggy performance in the first three rounds. If we take out the 
three games in question, and only count the games played by Deep 
Thought 0.01, then the performance for its fifteen games (six from 
the Fredkin Masters Open, and nine from the final rounds of the 
US Open) would be around 2592. Had Deep Thought 0.01 drawn 
the game against Lev, its fifteen-game performance would be 
around 2634. 

The US Open was a sobering experience. First, we had bugs to 
fix. Second, the Fredkin Intermediate Prize was obviously within 
reach, but the question was whether Hitech would beat us to it. 

Hitech was not doing much tournament-wise, so we had 
some time to regroup. We spent the next two months ironing out 
the bugs in Deep Thought 0.02. During the process, we found 
another major bug. Deep Thought was not generating en passant 3 
moves for the part of the search controlled by the microprogram. 
It had been puzzling us for a while that Deep Thought 0.01 was not 
playing better than ChipTest in head-to-head competition, despite 
having far more chess knowledge than ChipTest. The en passant 
bug might be the main culprit. 

Deep Thought 0.02 played in two more tournaments in 
October 1988, the American Open from October 8 to 10 in Los 
Angeles, California, and the US Class Championship from October 
28 to 30 in Somerset, New Jersey. Stuart Cracraft and Jim Gillogly 
graciously served as the operators for us in the American Open. 
Andreas traveled to New Jersey where Deep Thought defeated two 
more International Masters. After the US Class Championship, 
Deep Thought had played twenty-nine rated games. The overall 
performance for the twenty-nine games, including the three games 
marred by the mating bug, was 2509. The best twenty-five-game 
performance, again including the three bad games, was 2519. In 

3 For people who do not play chess, here is an explanation of the en passant rule. 
In chess, all pawns start on the second rank of the board. Before Renaissance, 
pawns could only move one square forward at a time. To speed up the game, dur- 
ing the Renaissance, the rules were changed to allow the pawn to move two 
squares forward on its first move from the second rank. This, however, allowed 
the pawn to escape being captured by opponent's pawn(s) on the adjacent files 
that normally could capture the pawn on the third rank. The en passant (in pass- 
ing) rule was added. Immediately after a two-square pawn move, the opponent 
can, with an adjacent pawn, capture the pawn just moved as if it were a pawn 
on the third rank. 



The Race for First Machine Grandmaster 81 

other words, we could have claimed the Fredkin Intermediate Prize 
right after the US Class Championship. However, we had no idea 
that Deep Thought had already achieved the milestone; we only 
knew that it was close, based on the wall-chart ratings, which were 
a few months out of date. Not knowing that we had already won 
the race turned out to be a good thing. The following month Deep 
Thought was about to shock the chess world. 

Given that ChipTest was slightly ahead of Deep Thought 0.01 up 
to the US Open in head-to-head competitions, it was entirely pos- 
sible that it could have claimed the Fredkin Intermediate Prize in 
1987, if only we had let it play in human tournaments. But then, 
the development of Deep Thought would have suffered. 

Ahead of a World Champion 

In January 1988, at a press conference in Paris, World Chess 
Champion Garry Kasparov was asked whether a computer could 
defeat a Grandmaster in tournament play before the year 2000. He 
replied, "No way, and if any Grandmaster has difficulties playing 
computers, I would be happy to provide my advice." Garry had no 
idea what was about to happen in the very same year. Nor did he 
have any inkling as to what kind of difficulties he himself would 
be facing nine years later in 1997, three years before the year 2000. 

At the time Garry made the prediction, it seemed to be a safe 
one. Computers had indeed beaten Grandmasters in blitz (five- 
minute) games more than a decade earlier, with Grandmaster 
Michael Stean and Grandmaster Robert Huebner as the victims. 
But under tournament conditions, computers had never drawn 
even a single game against a Grandmaster. While ChipTest was 
already in its full strength, it was unknown outside the computer 
chess community. As a matter of fact, not even the team really 
knew how strong ChipTest was, since the program did not have a 
rating until the Fredkin Masters Open. Cray Blitz was the World 
Computer Chess Champion, but International Master David Levy 
humiliated it in a 1984 match, winning all four games. The high- 
est rated chess computer, Hitech, for years only hovered between 
2350 and 2400 on the US scale. Grandmasters were typically rated 
above 2550 on the same scale. Surely, Garry's prediction had the 
ring of truth to it. 

In 1988, the Hall of Fame Chess Festival tournament was held 
in Canton, Ohio, from November 4 to 6. Deep Thought played 



82 Chapter 5 

with Mike Browne as the operator. Deep Thought had a good tour- 
nament, winning the first four rounds, including one over another 
International Master. In the fifth round, Deep Thought played Igor 
Ivanov for the second time in its short career. Igor had White and 
this time did not do any magazine reading! The game ended in a 
fighting draw. Deep Thought had its first tournament win, as it 
tied for first with Igor. 

The next weekend, Deep Thought played in its first ACM 
Championship in Orlando, Florida. Deep Thought won the title on 
tiebreak over Fidelity Challenger, which it drew in the first round. 
In this tournament Deep Thought played Hitech in public for the 
first time and won. (Deep Thought played Hitech three more times 
in computer chess events, lost one of them and won the other two. 
The winning ratio of 3 to 1 between the two programs is consistent 
with the roughly 200 rating points separating the programs.) Deep 
Thought was the clear winner or the co-winner of all the Computer 
Chess Championships that it played in. But luck played a big part 
in the results. Given the Swiss tournament format of these 
Championships, Deep Thought, despite being the heavy favorite, 
had only about fifty-percent chance of winning each individual 
event. 

Between November 24 and 27, the Software Toolworks 
Championship was held in Long Beach, California. With a total 
prize fund of $130,000, it was one of the top three US chess tour- 
naments in 1988. Besides several strong US Grandmasters and 
International Masters, there were three invited legends: the former 
World Champion Mikhail Tal, the former World Championship 
Candidate "Great Dane" Bent Larsen, and the former World 
Championship Candidate Samuel Reshevsky. Both Mikhail and 
Bent were still active and ranked high on the world-rating list, six- 
teenth and forty-second in the world, respectively. 

Again Stuart Cracraft, and Jim Gillogly offered their invalu- 
able help to be the remote operators. Jim, who happened to know 
one of the founders of Software Toolworks (known as Mindscape 
these days), smoothed the problems with the phone line with the 
help of Toolworks' personnel. Professor Raj Reddy, then a senior 
professor in the Computer Science Department at Carnegie 
Mellon 4 , offered to foot Deep Thought's bill for the tournament. 

4 Raj later became the Dean of the School of Computer Science at Carnegie Mellon 
University. 



The Race for First Machine Grandmaster 83 

(It would be many years before I knew that Raj was also instru- 
mental in setting up the Fredkin Prize in the first place.) By the 
time of the tournament, we knew that Deep Thought might have 
already won the Fredkin Intermediate Prize but there was some 
ambiguity on how the "performance" should be calculated. Deep 
Thought would have fallen short if the mathematically incorrect 
formula were to be used. We would have needed two more good 
games in that case. 

While Deep Thought had played in tournaments for half a 
year, the chess world was mostly unaware of its existence. The 
tournament director did a double take when he heard Deep 
Thought's rating and phoned the US Chess Federation to make 
sure that it was correct. An old acquaintance of mine, Sanjoy 
Mahajan, was playing in the tournament and chanced on IM 
Patrick Wolff, who later became a Grandmaster and US Champion. 
Patrick asked what Deep Thought's rating was at the beginning of 
the tournament. Sanjoy said it was 2510, to which Patrick said, 
"No way is it that strong." Patrick was to change his tune consid- 
erably as the tournament progressed. 

Deep Thought won the first two games without much trou- 
ble. As far as the Fredkin Intermediate Prize was concerned, we 
were home and dry. Stuart was the operator in the second game. 
The normal etiquette at the chess table is that one does not 
announce that one is about to checkmate the opponent. The old 
masters from the last century were known to announce check- 
mates, but it is considered bad form today. At the end of the game, 
Deep Thought calculated that it could checkmate its opponent in 
19 moves, that is, 19 moves by Deep Thought plus 18 replies by 
the opponent, for a total of 37 plies. Stuart was so excited that he 
told the opponent. The opponent resigned promptly. The posi- 
tion was quite lost even without the immediate mating 
possibility. 

The next day, we got word from California that our third 
round opponent was none other than the "Great Dane", Bent 
Larsen himself. Bent was one of the most successful Grandmasters 
in the postwar era. He has been a World Championship Candidate, 
not just once, but in four different World Championship cycles. In 
"the USSR vs. the Rest of the World" team match held in 1970, he 
played the first board for the Rest of the World, ahead of Bobby 
Fischer. At his peak, he had been as high as number 2 in the world. 
After the thrill of winning the Fredkin Intermediate Prize, it was 



84 Chapter 5 

time for a reality check. Bent would have White. I was hoping for 
the best and bracing myself for the worst. I was also a little bit envi- 
ous of Stuart and Jim. One does not get to see, let alone face, a 
legend every day. 

Bent was not known as a conformist, and he played unusual 
chess even when facing a computer. A few years before this game, 
I had read a biography that mentioned his predilection for push- 
ing the rook pawns, the two pawns on the outside edges of the 
chessboard. Deep Thought had showed a similar tendency in its 
first tournament, but I am quite sure that it had entirely different 
reasoning from Bent's. True to Bent's form, in this game, he 
allowed Deep Thought to have an advanced center pawn, and he 
did push both of his rook pawns before either of his center pawns. 
It is difficult to judge the position even after all these years. A 
human commentator asserted that Bent had an edge after the 
opening phase of the game. Deep Thought was quite happy about 
its position. It had an advanced pawn that cramped Bent's posi- 
tion. In return, he had considerable pressure on the advanced 
pawn. Against a human player, it might indeed be a decent posi- 
tion for Bent, as human players might have problems handling the 
pressure on the pawn. A computer would just calmly calculate all 
the variations, making sure that the pawn is not lost or at least not 
without exacting comparable compensation. As the game pro- 
gressed, Deep Thought's evaluation climbed up in its favor. I was 
beginning to think the unthinkable. Could Deep Thought indeed 
have the better position? 

Recent analysis, with a more powerful chess computer (Deep 
Blue Jr), showed that Deep Thought could have played a combina- 
tion and possibly won the game outright on move seventeen 5 . 
Deep Thought did not play the combination, but the position was 
probably still favorable. The human analysts were universally of 
the opinion that Bent had the better position throughout the early 
part of the game. My observation over the years is that only really 
good analysts can stay objective when analyzing human vs. com- 
puter games. 

Slowly, Bent improved the scope of his pieces. The position 
was very dangerous for both players, probably more so for him. 
Then on move twenty-six, Bent made a slight error. He moved his 
king and it happened to block the retreating square for his bishop. 

5 See Appendix B for the game score and the potential combination. 



The Race for First Machine Grandmaster 85 

Deep Thought threatened to trap the bishop with its next move. 
Bent could have just moved his king again to create an escape 
square for the bishop, although with a loss of time. When inter- 
viewed by IM David Levy a few weeks later, Bent mentioned the 
king move as "one very bad move". In reality, the bad move was 
the next one. Not willing to concede the mistake over the board, 
he lashed out at Deep Thought's king by pushing one of his pawns 
forward. The position was close to losing for him, and after his 
next move, Deep Thought calculated a forced win. Fifteen moves 
later, Bent resigned and became the first Grandmaster to lose to a 
computer under regular time control in a tournament. In 
Pittsburgh, we looked at each other in disbelief. We were on an 
emotional high. But we would come down to earth within a few 
hours. 

In the fourth round, Deep Thought played against six-time US 
Champion Grandmaster Walter Browne. Walter had played against 
computers in public before. In 1978, he lost a game against the 
Northwestern chess program, Chess 4.6, but he was also playing 
against forty-three other opponents at the same time! This time, 
however, he was playing against Deep Thought with undivided 
attention. There is an old saying in chess, "Sometimes you give 
them lessons, and sometimes they give you lessons." In this game 
Walter gave Deep Thought a good lesson. Deep Thought at the 
time did not understand the value of a bishop controlling an 
"open diagonal" or the value of a "permanent pin" 6 , and Walter 
exploited the weaknesses beautifully and won the game fairly eas- 
ily. Deep Thought put up quite a struggle, but it was not enough. 

Deep Thought drew with an IM for the second time ever 7 in 
the fifth round game against Vince McCambridge, and then won 
the next game against a strong master. In the seventh round, it was 
paired against new IM Alex Fishbein. A couple of years ago, at the 
ACM Championship, IM Danny Kopec brought in an article he 
wrote about the ending rook and bishop vs. rook. According to 
chess endgame books, this is a typically drawn ending. Danny 
Kopec's article, however, mentioned that even Grandmasters fre- 

6 An open diagonal is a diagonal that is no longer impeded by pawns of either 
side. A permanent pin is an unbreakable pin, typically formed when a piece is 
pinned against its own king by an opponent's bishop. 

7 Up to this point, Deep Thought had won all its games against IMs, with the 
exception of one draw against Igor Ivanov. 



86 Chapter 5 

quently made game-losing mistakes in this ending. Since ChipTest 
had no problem finding the mating sequence from Danny's sam- 
ple winning positions, we made a conscious decision to allow 
ChipTest, and later, Deep Thought, to play into the ending. Prior 
to the game against Alex Fishbein, Deep Thought had won twice 
from the stronger side of this ending against players of expert level 
and this decision seemed to be sound. I greatly regretted the deci- 
sion when Deep Thought allowed Alex to convert from a clearly 
lost position into the weak side of rook and bishop vs. rook end- 
ing. Theoretically, the position on the chessboard was a draw with 
perfect play. 

Sanjoy Mahajan talked with senior master Adam Leif and IM 
Patrick Wolff, who both thought Alex should know how to play it. 
"He's just the kind of player who'd know it," said Patrick. For a 
while, Alex was maintaining the draw. Since the final round was 
coming up, the tournament directors adjourned the game to be 
played after the last round game. There was an interesting debate 




The Deep Thought team. 

From left to right: Munay Campbell, Feng-hsiung Hsu, Thomas Anantharaman, 
Mike Browne, and Andreas Nowatzyk. In the foreground are the Deep Thought 

board itself and the CDC plaque (plaque donated by the supercomputer 

company CDC for the ACM Championship). Taken after Deep Thought won the 

Fredkin Intermediate Prize for its Grandmaster level performance in 1 988. 



The Race for First Machine Grandmaster 87 

among the tournament directors about how Deep Thought should 
be paired in the last round. If the game had been paired as a win 
for Deep Thought, our last round opponent would have been 
British Grandmaster Anthony Miles. Instead, the game was paired 
as a draw, and we ended up playing IM Jeremy Silman who hap- 
pened to be Stuart Cracraft's chess teacher, and Stuart was nervous 
facing his own teacher. Of course, Deep Thought did not know 
who Jeremy Silman was. Jeremy managed to steer the game into an 
even position, but then lost in fifty-seven moves in a difficult end- 
ing. 

The adjourned game against Alex Fishbein was played after 
both players had finished their last rounds. Alex lost his to GM 
Anthony Miles, who was leading the tournament with 6.5 points 
out of 8. Deep Thought was the only player that could catch up 
with Anthony. Alex's sealed move turned out to be a losing one 




With Ken Thompson and David Slate. 

Standing from left: Ken Thompson (Turing Award winner 

for UNIX and one of the authors of Belle) 

and David Slate (one of the authors of Chess 4.X). 

Taken at the 1989 WCCC (World Computer Chess Championship). 



88 Chapter 5 




Claude Shannon presents the author with a horse's head. 

Taken at the 1989 WCCC 

and Deep Thought quickly found the winning sequence. This time 
Stuart did not announce the mate, and the game was played out to 
move eighty before Alex conceded. 

GM Anthony Miles and Deep Thought were the co- winners of 
the tournament, ahead of six other Grandmasters, including the 
former World Champion Mikhail Tal. This was the first time that 
any computer finished first in any tournament ahead of 
Grandmasters, let alone a former World Champion. Since Deep 
Thought was not eligible for prize money, Anthony Miles was 
awarded sole possession of first prize. 

The race for the Fredkin Intermediate Prize was officially over. 
Deep Thought performed at 2776 on the US scale for the tourna- 
ment, slightly behind the Hall of Fame result, which was 2790. In 
1988, Deep Thought played forty-two rated games, and the per- 
formance over all forty-two 8 , including the ones played with 
serious bugs in the US Open, was 2598. The best twenty-five-game 
performance over the period was 2655 or 155 points higher than 

8 It is a strange coincidence that 42 appeared repeatedly in this chapter, first as 
Deep Thought's answer to the question of "Life, the Universe, and Everything", 
then as the world ranking of Bent Larsen, and the number of rated games by 
Deep Thought in 1988. 



The Race for First Machine Grandmaster 89 

the performance rating requirement for the Fredkin Intermediate 
Prize. 

A few months after the Software Toolworks competition, Deep 
Thought became a World Champion in its own right in 
Edmonton, Canada, at the World Computer Chess Championship 
tournament that was held every three years. Computer chess com- 
petitions, however, were becoming less interesting to us. Our full 
attention turned to the ultimate quest of beating the human World 
Chess Champion. 

As a final note, in 1989 the organizers of the Software 
Toolworks Championship arranged for Deep Thought to play GM 
Anthony Miles, in an exhibition game as the unofficial tiebreaker 
for the 1988 Championship. Deep Thought won the game. 



ft 



CHAPTER 6 



"Knock, Knock. Who's There?" 



The Man Who Likes to BE Your Friend 

Frederic Friedel is a German freelance journalist, that is when he is 
not too busy working for the two computer chess-related compa- 
nies that he co-founded. One of his companies publishes a 
computer chess magazine Computer Schach und Spiele, which 
reviews and recommends chess computers and chess software, 
while the other company, ChessBase, produces chess software. 
There is a potentially serious conflict of interest. But Frederic is not 
someone who would be stopped by a mere conflict of interest, and 
he is quite open about it. 

Andreas Nowatzyk knew Frederic from his time at the 
University of Hamburg. He had met him a couple of times through 
some friends who were active in computer chess. Andreas 
described Frederic as someone who drove fast cars, wore designer 
suits, and hung out with famous people. The "hung out with 
famous people" part is still correct, but the Frederic I know today 
is a family man with a loving wife and two bright kids. 

Frederic had organized some fairly large computer chess 
events on German soil. One of them was the 1979 TV match 
between 1M David Levy and Chess 4.8. Either Frederic was very 
persuasive or David was very brave. A gigantic industrial robot 
made the moves for chess 4.8 over the chessboard. The robot could 
easily have made David a very famous but damaged IM, if there 
had been a power flicker or two. 

Frederic was a friendly sort of guy who could be very helpful, 
especially when the help was mutual. Since both of Frederic's com- 
panies are in computer chess, he socializes with many of the top 
chess players, usually offering them free chess software and data- 

90 



"Knock, Knock. Who's There?" 91 

bases of chess games. He became a close friend of Garry Kasparov 
very early, possibly even before Garry was crowned the World 
Chess Champion. Garry was said to be instrumental in the forma- 
tion of ChessBase itself. Apparently, the ChessBase software, which 
is a database program for chess games, was created to satisfy 
Garry's needs and many of the features of the program were based 
on Garry's specifications. 

In late July 1988, Andreas received at our office a phone call 
from Frederic. Afterwards, he sent an e-mail to the rest of the team 
summarizing the conversation. Frederic was inviting us to play in 
an exhibition event in Germany. I was quite surprised. Although 
ChipTest had won the 1987 ACM Championship, Hitech was the 
better known machine. In fact, Eurtpe Echecs, a French chess mag- 
azine had a report after the Championship that said Hitech was 
beating ChipTest by the ratio of 2 to 1. The Europe Echecs report 
probably misquoted or cited a wrong source, as the tally from 
ChipTest's log showed something completely different. Anyway, to 
the outside world, Hitech was the uncrowned king. The results 
from the Fredkin Masters Open for Deep Thought and ChipTest 
dispelled the myth, but it was unclear as to whether they were well 
known outside of Carnegie Mellon. How did Frederic find out 
about us? 

It turned out that Frederic had befriended Ken Thompson 
back in Belle's heyday. Ken was spending a sabbatical year from 
Bell Labs in Sydney, Australia. Since Ken was on my thesis com- 
mittee, I kept him updated about what was happening with 
ChipTest and Deep Thought, and he apparently told Frederic about 
us. Sometimes, it pays to know the right people. 

The German invitation was part of a birthday celebration in 
the fall for Conrad Zuse, Germany's most prominent computer 
pioneer. Conrad had made a prediction fifty years earlier that a 
computer would defeat the chess champion within fifty years. A 
big event was planned to match the best computer against Garry 
Kasparov. Frederic was organizing it, and wanted Deep Thought to 
be the computer. It was definitely too early to face Garry but it was 
too good a chance to pass up. Anyway, it would be an exciting 
thing to do. 

Since we were busy chasing after the Fredkin Intermediate 
Prize, we did not really keep track of what was happening with the 
invitation. Andreas maintained contact with Frederic and kept the 
rest of the team informed. Frederic also offered to provide Deep 



92 Chapter 6 

Thought with game databases, and the offer was gladly accepted. 
By mid-September 1988, there were a few new developments. A TV 
scheduling conflict meant that the match was delayed until 
January. Another complication was that IBM Germany was the 
sponsor and was pressing hard to run Deep Thought on an IBM 
host computer. Deep Thought could not run on an IBM host at the 
time so the funding for the match was in doubt. A week later, we 
were informed that the match was off. Thus ended the first 
attempt to arrange a match between Garry and Deep Thought. 

Frederic kept in touch with the team over the years. Since he was 
also close to Garry, we got to hear some of the Champion's opin- 
ions about chess computers. The following excerpts came from a 
letter he sent me in 1989: 

Beginning 1988: "Computers will never beat Grandmasters." 
Mid 1988: "Computers will never beat strong Grandmasters." 
End of 1988: "Okay, yes, Bent is a strong Grandmaster. But a 
computer will never be able to beat Karpov or me." 

Murray Campbell had an interesting observation in regard to 
this last statement. Murray had been in the computer chess field 
far longer than I had. Over the years, he had observed that when 
computers were much weaker, chess masters would say that com- 
puters would never beat chess experts. Of course, experts are one 
level below masters in playing strength. As time progressed, the 
same pattern repeated with players of greater strength. Anatoly 
Karpov happened to be the second strongest player of the world 
behind Garry Kasparov when Garry made his prediction. 

The IBM Connection 

IBM Germany called off the German match with Garry in 
September 1988. Meanwhile, IBM Research in the US had been 
talking with us for some time about the possibility of continuing 
our research at IBM. As far as I know, IBM Germany had no idea 
that we might be moving to IBM after graduation. 

The first hint of IBM interest came from Kai-Fu Lee, a 
Carnegie Mellon graduate student working in the speech recog- 
nition area. Kai-Fu talked to some IBM speech recognition 
people about what the Deep Thought team had done, and got 
the feedback that IBM might be interested in supporting our 



"Knock, Knock. Who's There?" 93 

research. One of the IBM people that Kai-Fu talked to was Peter 
Brown. 

Peter and I had been in the same 1982 incoming class at 
Carnegie Mellon. We knew each other, but were not really close. I 
bumped into Peter in the Computer Science lounge a few weeks 
before the Fredkin Masters Open, and mentioned what I was 
doing. Peter was asking a lot of questions, but I had no idea that 
he was thinking about getting IBM to hire us. 

Years later, Peter described how he got the upper management 
in IBM interested in continuing the project inside IBM. It was 
around the time of the Superbowl season. Peter happened to be in 
the men's room with Abe Peled, the IBM Research Vice President of 
Computer Science. They started talking about how expensive the 
Superbowl TV commercial spots were. Peter suggested that he 
knew a way to gain much greater publicity at far lower cost. He 
knew this group of graduate students working on computer chess 
at Carnegie Mellon, and he believed that this team would create 
the first chess machine to beat the World Champion. Given the 
historical significance of the quest, and the latent public interest, 
IBM could stand to gain huge advertising value from the endeav- 
or. Oh, yes, members of the team were world-class people that IBM 
would be interested in hiring anyway. Abe got interested and asked 
Peter to look into it. 

John Cocke, the now retired IBM fellow and Turing Award 
winner who pioneered the idea of RISC (Reduced Instruction Set 
Computer) among other things, was also instrumental in getting 
the team hired. John is a good friend of Professor Edward Fredkin, 
who donated the money to set up the Fredkin Prize for computer 
chess. John was quite excited about the possibility of the top com- 
puter chess team coming to IBM, and his excitement was 
contagious. By early November 1988, a couple of IBM managers 
were interested and we were asked to send in applications. This 
was before we made the splash at the Software Toolworks 
Championship. 

Thomas, Murray and I visited IBM Research separately and 
gave our sales pitch. The idea was to integrate an entire chess 
machine onto a single silicon chip, and then use a massive num- 
ber of the chips to build the "ultimate" chess machine — a chess 
machine that could beat the World Champion. I would be the one 
doing the chip work, while Thomas and Murray would develop the 
software first on an intermediate machine, and then port it to the 



94 Chapter 6 

final machine. By early March 1989, both Thomas and I were 
offered regular positions at IBM Research to work on the chess 
machine. There were some complications with Murray, and it took 
another month or so before he was offered a post-doc position. His 
position later became a regular one. 

A few weeks before the notice from IBM, I received an e-mail 
from the manager of the Computer Science Laboratory at Xerox 
PARC (Palo Alto Research Center), asking me to consider PARC. 
Xerox PARC was a legendary place for computer science research. 
Some of the most exciting computer science research in the 1970s 
and 1980s came from there. For instance, the main idea behind the 
Apple Macintosh can be traced to the PARC lab. Some of the most 
innovative early VLSI design software also came from PARC. In 
addition, the lab atmosphere was much closer to what I was used 
to at Carnegie Mellon. PARC was also right in Silicon Valley. I had 
been there a few times, and really enjoyed the different atmos- 
phere, not to mention the weather! The fact that Silicon Valley was 
much closer to Taiwan was a bonus. I quickly sent in my vita to 
PARC and received a positive initial response. 

When the IBM offer came in, I was still considering PARC as 
an alternative. Both Thomas and Murray accepted the IBM offer 
immediately, while I was still waiting to find out whether PARC 
provided a better option. I was in a fairly awkward position. If I 
were to decline the IBM offer, Thomas and Murray would need to 
find a new project at IBM, and I did not want to put them in a dif- 
ficult situation, especially after what we had gone through 
together. But I had also promised the PARC people to give them a 
fair look. Furthermore, I liked the idea of going to California. "Go 
west, young man." I always thought I would end up in Silicon 
Valley. A lot of my college classmates were there, and several of my 
good friends at Carnegie Mellon were also heading that way. I 
decided to delay the decision until I found out what PARC had to 
offer. 

Xerox PARC was a very interesting place, especially to some- 
one who worked in the VLSI area. However, I did have some 
reservations. Would Xerox, which was not really in the computer 
business, be willing to support the effort long term? I also did not 
know what decision I would make if Xerox offered to support the 
project, but hired only one of us. 

Part of me welcomed the idea of working alone again, at least 
for a while. The team had been working in an assembly line fash- 



"Knock, Knock. Who's There?" 95 

ion, with me completing the hardware first, and the others finish- 
ing the software afterwards. Theoretically, the hardware and the 
software can be developed simultaneously, but that is only possi- 
ble when the hardware is well defined, which was certainly not the 
case with the new chess chip. Also, the new chip could easily take 
a few years to design, while the software could be completed in say, 
6 months to a year. IBM's hiring of the team had the advantage of 
keeping the team intact, as well as providing IBM immediate exter- 
nal visibility. But as far as the project schedule was concerned, 
there was relatively little to be gained; in fact, the need to provide 
external visibility could easily delay the project, as we would then 
have to build an intermediate machine to ensure that we remain 
on top of the computer chess world. Working alone, 1 would have 
been able to go directly to the new chess chip, which was what 1 
really wanted to work on. 

Anyway, I did not tell the people at PARC that I would con- 
sider the option of working alone. Perhaps, subconsciously, I did 
not want to face that difficult choice should it arise. I did mention 
what IBM was willing to offer. In the end, PARC decided that they 
could not match IBM's offer to hire a full team, and I informed IBM 
Research of my acceptance. 

The Poor Lieutenant Colonel at DARPA 

After the arguments over the invitation from California in 1987, I 
avoided dealing with Dr Berliner as much as possible. He was still 
on my thesis committee of course, but it made no sense to rock the 
boat given the less than ideal relationship we had. Allen Newell, 
who was one of the founders of the Department and also one of 
the pioneers of computer chess, replaced Dr Berliner on my thesis 
committee in April 1989. In retrospect I wondered if the reason for 
Dr Berliner leaving the committee were more to do with an inves- 
tigative reporter than with the man himself. 

Deep Thought's defeat of GM Bent Larsen, and tying for first 
in the 1988 Software Toolworks Championship, were big news in 
the chess world. At the same time of the Toolworks Championship, 
the twenty-eighth Chess Olympiad was being held in Thessaloniki, 
where nearly all of the world's top chess players competed for their 
national teams. When the news of Deep Thought's win in Long 
Beach was wired to Thessaloniki, it became the talk of the day. 
Probably most of the top chess players had never heard of Deep 
Thought until then. 



96 Chapter 6 

It took a while before the general press picked up on the news. 
On December 22, a month after the tournament, the front page of 
the International Herald Tribune carried the headline "A Computer 
First: 'Deep Thought' Stuns the Chess World". None of us on the 
team had any idea that we had made the front page of an interna- 
tional newspaper. As far as we knew, we did not make the front 
page of any national newspaper in the US, and certainly we were 
not expecting to see the news showing up in an international 
newspaper. This headline caught the attention of a reporter in 
London who decided to start an "investigation". 

We learned about the reporter's interest in the project from Dr 
Berliner, who called us to say that a reporter was interested in playing 
a demo game with Deep Thought. Dr Berliner added that the reporter 
was from a well-respected magazine. Since we were starting to look 
for jobs and could use the publicity, we readily agreed to set up the 
demo. At the agreed time, the reporter called and introduced himself 
as Dominic Lawson from the British magazine The Spectator. I found 
out later that he was the Deputy Editor of the magazine (he is now 
Editor of Tlie Sunday Telegraph newspaper in Britain). Dominic lost 
the game and we thought that was the end of the story. He made no 
attempt to interview any member of the Deep Thought team and 
seemed only to be interested in playing the game. 

Then on 27 February 1989, someone showed me the column 
of Charles Krauthammer, syndicated from the Washington Post, in 
the Pittsburgh Post-Gazette. It was a very well written article about 
Deep Thought's success in Long Beach and its implications, and 
quoted The Spectator as the source for some of the material. 

One of the libraries at Carnegie Mellon had a subscription to 
The Spectator and I found the article quoted by Charles 
Krauthammer. It was not just an article, but the cover story of that 
issue. On the cover, an airplane, with the fuselage in the shape of 
the chess piece bishop and four engines in the shape of chess 
pawns, was flying over a checkered field. An American flag was 
prominently displayed. The cover headline read "Playing to Win" 
and the smaller headline read "Dominic Lawson reveals the 
Pentagon's war moves". I frowned, "Andreas is not going to like 
this." Andreas had a sympathetic view for Germany's Green Party, 
which was into environmental causes and pacifism. Supporting 
any sort of war effort would be completely counter to his beliefs. 
For that matter, none of the other members of the team was 
thrilled about the connections that were made. 



"Knock, Knock. Who's There?" 97 

In one sense, Dominic's article was hilarious. The Spectator 
appeared to be a magazine that took itself seriously, but the article 
seemed to me to be pretty groundless. The subtitle inside the mag- 
azine read "Dominic Lawson investigates the Deep Thought 
project, and discovers that the Pentagon is harnessing chess to the 
arms race". Okay, I did not know — and I still do not know — 
whether the Pentagon was harnessing chess to the arms race. But 1 
do know that Deep Thought was not started because someone in 
the Pentagon wanted it done. 

Once I started reading the article, factual errors kept on pop- 
ping up, and that was just for the parts that I knew for sure were 
wrong. One section of the article read, "The Deep Thought project, 
and other computer chess developments at Carnegie-Mellon, have 
been paid for in their entirety by DARPA, the Pentagon's Defense 
Advanced Research Projects Agency". So, the money we paid out of 
our student pockets belonged to DARPA as well! 

Then I read what was the most amusing part of the article, "i 
asked Lieutenant Colonel Robert Simpson, DARPA's head of fund- 
ing for expert systems research — which includes the Deep Thought 
project — to give me one clear military application of Deep 
Thought." First, Deep Thought was not an "expert system", which 
means something quite specific in the research field of Artificial 
Intelligence. Several members of the team would probably consid- 
er it an insult to call Deep Thought an expert system. In fact, at 
least two members of the team had used the word bullshit to 
describe "expert systems", or for that matter, Artificial Intelligence. 
Second, the Deep Thought research was supported in part by the 
VLSI project, and none of the money came from the DARPA fund- 
ing for expert systems research. Lieutenant Colonel Robert 
Simpson had to my knowledge never received a single scrap of 
paper describing the Deep Thought project. 

Just imagine the situation that this Lieutenant Colonel was 
placed in. A reporter calls you up and asks you about a project that 
you know nothing about. You have to say something. Well, 
Lieutenant Colonel Robert Simpson had a quick mind. 

" 'Navigation in a battlefield situation,' he replied promptly. 'A 
machine like this, programmed with knowledge of the terrain a 
pilot is flying through, can digitize all the various route choices, 
explore them, and choose the optimum route. That is exactly what 
Deep Thought is doing.'" Interesting, I hadn't thought that Mr. 
Kasparov would make a nice cruise missile. 



98 Chapter 6 

Then I came to something quite annoying. "The head of the 
Deep Thought project and of the Carnegie-Mellon computer sci- 
ence department, Professor Hans Berliner, warned me that in about 
three years his team would have produced a son of Deep Thought, 
which would be stronger at chess than the world champion, Garry 
Kasparov. 'It's a pity, really,' he sighed. 'Just as chess players were 
beginning to make a decent living.'" Dr Berliner was neither the 
head of the Deep Thought project nor the head of the department, 
and he was a competitor. What had gone on in Dominic's inter- 
view with Dr Berliner, who had had absolutely nothing to do with 
Deep Thought? 

When I calmed down, I realized that the reality might be that 
Dominic had a good story in mind, and may not have had the 
time to check all the details. Dr Berliner had after all erroneously 
been named as leader of the project in other media coverage of 
Deep Thought. 

Anyway, after the rest of the team had seen the article, we 
decided to send a letter to The Spectator. I also sent an e-mail to Dr 
Berliner asking him to please make sure that other reporters would 
not make the same mistake. We were still looking for jobs, and the 
wrong attribution could affect our chances. One thing led to 
another, and a month later, Professor Allen Newell replaced Dr 
Berliner on my thesis committee. 

A few weeks after we mailed the letter to The Spectator, a mas- 
sively modified letter appeared in their letters section. 

Sir: The article, 'The Pentagon Plays Chess' (28 January), con- 
tains several gross misconceptions and factual errors. 

(This is different from what we actually wrote, but the essence 
is the same.) 

The Deep Thought project is not entirely funded by DARPA, 
the Department of Defense, or the Pentagon directly. The 
support includes industrial funding, National Science 
Foundation funds, and, to some degree, Department of 
Defense money. 

This is drastically different from what we wrote, which was 
"The Deep Thought project is not funded by DARPA/DOD, or the 
Pentagon directly, and certainly not in its entirety. The Deep 
Thought effort was a recreational exercise where several graduate 
students applied what they learned about VLSI circuit and hard- 



"Knock, Knock. Who's There?" 99 

ware design. The first result was ChipTest, which we built entirely 
out of scraps from other projects in spare time. After ChipTest's 
success in the 1987 ACM Computer Chess Championship, one of 
our advisors, Prof. H. T. Kung, helped us with a modest $5000 out 
of his discretionary fund, which was used to construct the current 
Deep Thought. This money is not directly related to any official 
project, but is part of normal operational expenses of typical uni- 
versities. As such, it includes industrial funding, tuition, NSF 
(National Science Foundation) funds, donations, and to some 
degree, DOD money. In addition, some of the various expenses for 
Deep Thought came right out of our own pockets." 
The published letter continued. 

We very much doubt that Lieutenant Colonel Robert Simpson 
was aware of the nature of Deep Thought which is com- 
pletely unrelated to expert systems research. Deep Thought 
is a special purpose engine for playing chess and chess 
only. The principal algorithms used by Deep Thought had 
been known for decades, and the structure of Deep 
Thought has no utility for battlefield navigation whatso- 
ever. To quote Ken Thompson of Bell Labs, "the only 
military application for Belle [a chess machine like Deep 
Thought] is to drop it from an airplane and kill someone." 

This paragraph was given in its entirety. 

The following did not show up in The Spectator. "Dr Hans 
Berliner is a senior research faculty member of the CMU Computer 
Science Department, and heads the research project on the Hitech 
chess machine. He is not head of the Computer Science 
Department, who is Professor Nico Habermann, nor is he involved 
with the Deep Thought effort in any way. 

Mr. Lawson, in his 'investigation' of the Deep Thought proj- 
ect, managed the unusual feat of interviewing not a single member 
of the Deep Thought team." 

The letter must have been too long. 

From Chess to Wall Street 

Once I decided to go after the World Chess Champion, I read 
Grandmaster Robert Byrne's columns in the New York Times when- 
ever I could lay my hands on them. When ChipTest became 
operational, I would check out Robert's analysis with ChipTest. 



100 Chapter 6 

Sometimes ChipTest would come up with ideas that he missed, but 
frequently he would have some deep analysis that ChipTest had a 
hard time replicating. 

When ChipTest won the 1987 ACM Championship, Robert 
wrote an article about it but we did not really have contact with 
him until after the 1988 ACM Championship where Deep Thought 
defeated Hitech. Mike Valvo, the tournament director, was provid- 
ing Robert with background information about Deep Thought. 
Since many people did not realize that Dr Berliner was unrelated to 
Deep Thought, we asked Mike Valvo to help clarify the situation. 
It was our first indirect contact with Robert. 

While Deep Thought was never an Artificial Intelligence proj- 
ect, the AI community treated Deep Thought as one of their own, 
and we were invited to play an exhibition match with none other 
than GM Robert Byrne at the 1989 International Joint Conference 
on Artificial Intelligence. Neither player was on site. Deep Thought 
played from Pittsburgh, while Robert played from his home in New 
York State. The moves were relayed over the USA Today Sports 
Center, an on-line service similar to AOL or MSN today. Deep 
Thought won the game, but a few days later Robert returned the 
favor in a second game arranged by the Sports Center. Robert and 
Deep Thought played two more games with one win each. 

After these four games, the New York Times took an interest in 
the project, and Robert was asked to write an article for the Science 
Times section of the paper. He visited Carnegie Mellon for the arti- 
cle, and it was the first time we met him face to face. He looked like 
a distinguished old gentleman — the most frequent description 
used when people got to know him was "a sweet gentleman". We 
all liked him immediately. After we joined IBM, we found out that 
Robert was living close by. An avid tennis player as was Murray 
(before Murray's daughter was born), they would frequently play 
on weekends. Once in a while, Robert would drop in and play 
some games with our latest machine. Except for the Grandmasters 
who worked on the project, Robert probably knew more about our 
machines than any other Grandmaster. 

Robert's article in the Science Times was the cover story, span- 
ning two pages. In it, Andreas' work on the automatic tuning of 
the evaluation function was described. A CEO of a Wall Street stock 
and commodity trading firm saw the article and gave Andreas a 
call, offering him a high paying job. Out of curiosity, Andreas went 
to New York and met the Wall Street people. Wanting to impress, 



"Knock, Knock. Who's There?" 101 

the CEO took him to the stock exchange, but Andreas was wearing 
jeans, and jeans were not allowed on the floor. The CEO had to talk 
the security guard into looking the other way. The following floor 
trip was supposed to be for a peek only, but somehow it became a 
more detailed tour. The security cameras caught Andreas wearing 
jeans on the exchange floor, the security people descended, and an 
argument ensued. I don't know about Andreas, but if I were in his 
place, I would have had a hearty laugh once I was cleared from the 
floor. Wall Street was a different world. 

While the Wall Street proposal was interesting, Andreas 
turned it down. The Wall Street firm continued to sweeten the 
offer, but Andreas remained uninterested. 

In February 1990, Andreas accepted a job at Sun Microsystems 
in California, with the starting date in June 1990. He still had some 
writing to do for his thesis. By this time both Murray and I had 
already left for IBM. The Wall Street firm came back with an offer 
for Andreas to do some consulting work while he was finishing up 
the thesis. Thinking of the consulting work as a paid vacation from 
writing his thesis, Andreas went to New York for four months, with 
occasional trips back to Pittsburgh. At the time, the software used 
in the Wall Street firm was in pretty bad shape. It worked, but at 
the speed of a snail. For the piece of software that Andreas was 
working on, the total storage for the data was something like 10 
megabytes. The software always went back to the disk for the data, 
even though the disk access speed was more than a thousand times 
slower than the main memory access speed. Andreas rewrote the 
software to read in all the data into main memory, and voila, it 
now ran more than a thousand times faster. The Wall Street people 
were impressed. Andreas did some other amazing things in their 
eyes. To them, Andreas was like someone who could walk on water. 
Today, many real computer scientists work on Wall Street and the 
problem that Andreas found is unlikely to reappear. 

After Andreas went to work at Sun Microsystems, the Wall 
Street firm refused to give up on him and offered him a proposal 
that they thought he could not refuse. The proposal went some- 
thing like this: the firm would create a company in Silicon Valley 
with a hefty annual budget. Andreas would head the company and 
have full say on what the company would do, but half of his time 
had to be spent on continuing what he was doing for the firm. 
Andreas would be free to hire people who would also be given prof- 
it sharing deals and so on. Andreas asked Thomas Anantharaman 



102 Chapter 6 

and Mike Browne, as well as John Zsarnay, a staff engineer at the 
Engineering Lab, whether they were interested. They were. John 
went to California and started setting up the infrastructure. But 
then Andreas had second thoughts and backed out of the deal. The 
firm renegotiated the deals with Thomas and Mike, and John was 
moved back to the East Coast. John soon left and joined Rambus, 
a startup at the time in Silicon Valley. I was quite unaware of all 
this until much later, but the whole episode eventually would have 
an unexpected influence on the project at IBM. 

None of us on the Deep Thought team ever imagined that our 
work would have some link with Wall Street. The automatic tun- 
ing part of the project was an interesting topic in its own right. But 
the main reason for doing it came from the incident with the 
California invitation. Who would have thought that the invitation 
from California for a chess tournament would lead to some of us 
working for a Wall Street firm? Sometimes, life is full of the 
strangest twists and turns. 



Intermezzo 



CHAPTER 7 



First Date with History 



A Phone Call from New York 

After Deep Thought became famous in the chess world, phone 
calls started to come in from all sort of places. Usually, they were 
put through to our offices. One night, however, I got a call at home 
from some guy I had never heard of, who was asking lots of ques- 
tions. He was supposedly calling from somewhere in Minnesota, 
was not associated with any news organization, but was calling 
purely out of personal interest. Apparently, he found my phone 
number in the white page. He seemed harmless, but I did get nerv- 
ous talking to him. For several months until I left Carnegie Mellon 
for IBM, he would call every few weeks trying to find out what was 
new with Deep Thought. 

1 was therefore a little bit nervous when I got another unex- 
pected phone call at home. Although the voice did seem 
somewhat familiar, I was sure that it was not someone I knew. "Not 
another one," 1 thought. The caller introduced himself as Shelby 
Lyman. 

Before "the match of the century '" between Bobby Fischer 
and Boris Spassky in 1972, chess was not telecast in the USA. 
Shelby Lyman and his friends persuaded PBS (Public Broadcasting 
Systems) to do a TV show on the match, with Shelby being the 
main commentator. For a while, with the Fischer-Spassky match 
grabbing the attention of people on the streets, Shelby was an 
instant celebrity. The public's interest in chess waned after Bobby 
Fischer refused to defend his title under FIDE's conditions against 
his Russian challenger Anatoly Karpov. Shelby, however, main- 

1 Hey, hyperbole is free. 

105 



106 Chapter 7 

tained a good relationship with PBS and was able to get PBS to 
broadcast more TV shows covering the World Chess 
Championships. 

I started viewing some episodes of Shelby's show after decid- 
ing that I wanted to beat the World Champion. To tell the truth, 
Shelby's show was a little bit slow and too artificial for my taste. 
The show might have worked in 1972 when the interest in the 
Fischer vs. Spassky match was high, but with the MTV and video 
game generation, the show simply did not resonate. 

Shelby asked whether we would be interested in playing an 
exhibition match with Garry Kasparov, the World Chess 
Champion. Of course, the answer was yes. We were not really 
ready to play the world champion, but we did not want to miss the 
chance. I did have some doubt whether Shelby could make the 
match happen. The earlier exhibition match proposal from 
Germany was scrapped in the end, despite the close relationship 
between the potential organizer and Garry. 

At the time that Shelby called, I was designing a printed cir- 
cuit board version of Deep Thought. Professor Raj Reddy at 
Carnegie Mellon had taken an interest in the Deep Thought proj- 
ect, and offered to provide us with about $10,000 to make 
additional boards for a version of Deep Thought with more than 
two chess processors. The component cost had come down suffi- 
ciently that we could build four printed circuit boards with two 
chess processors on each board. In theory, for the proposed match 
we could have a new machine running that was up to five times as 
fast as Deep Thought (four new boards plus one existing board). 

Not knowing much about the mechanical design of printed 
circuit boards, I made a serious error. The Deep Thought boards 
were the size of a large pizza. For circuit boards of this size, the heat 
expansion from the soldering process can cause severe warping of 
the circuit boards. While I did place a board stiffener on the new 
printed circuit board to prevent warping during normal handling, 
it was not strong enough to prevent warping resulting from the 
soldering process. All four boards were quite visibly warped. 
Probably because of the warping, one of the four new boards never 
worked. It might have been possible to fix the bad board with a lot 
of work, but I had a thesis to defend and lots of packing to do 
before I left for IBM. Therefore, the maximum number of boards 
that we could really have for the match was four. 

After my initial positive response, Shelby came back with 



First Date with History 107 

more details for the proposed match. NYNEX, the Telephone 
Company for the New York area, would sponsor it. I mentioned to 
Shelby that some of us would be moving to IBM, but NYNEX 
apparently had no problem with the IBM connection. The match 
would be a short one, lasting only one day. Two games would be 
played, and the games would be played with 90 minutes to each 
player for the entire game. Each game would last at most three 
hours. Garry Kasparov's one-day appearance fee, not including 
expenses, for the match would be $10,000, which was supposedly 
heavily discounted. He viewed the match more as an experiment 
than as a real match and he was willing to give a discount. Anyway, 
$10,000 for one day of work is still a tidy sum. (In 1997, 1 was asked 
by IBM Asia Pacific to give a lecture and presentation tour for near- 
ly a month, stopping in seven countries. The total expenses, 
including the plane fare and hotel accommodation, came to slight- 
ly more than $10,000. Nowadays I believe that Garry's fee for a 
one-day appearance is something like $30,000, enough for me to 
take a three-month business trip around the world.) 

Murray joined IBM in September 1989 after finishing his stint 
as a research associate at Carnegie Mellon. I joined IBM a month 
later, after my thesis defense 2 . Thomas was still quite a way from 
completing his thesis and stayed behind at Carnegie Mellon. Prior 
to our leaving the university, IBM and Carnegie Mellon reached an 
agreement that we could take the three working Deep Thought 
printed circuit boards with us to IBM. Since Thomas was staying 
behind, I brought only one board with me. The other two new 
boards and the original Deep Thought board remained at Carnegie 
Mellon so that Thomas could complete the parallel code using all 
three boards. The three boards would be placed inside three work- 
stations, with one board in each. The workstations would then 
communicate over the departmental local area network at 
Carnegie Mellon and negotiate how the search would be run in 
parallel. This configuration was what we used at the Deep Thought 
vs. Kasparov match. It ran at two to three times the speed of the 
regular Deep Thought. It would not be competitive with Garry, but 
at least we could hope for a good showing. 



2 Well, if you were expecting any fireworks at my thesis defense, I have to say it 
went smoothly. The defense was open to the public and Dr Berliner attended it 
as a member of the public. While Dr Berliner might have disagreed with the 
directions that I was taking, he did not give me any trouble. 



108 Chapter 7 

But Not Karpov and Me 

I first heard the name Garry Kasparov from Mike Browne, my 
office mate at Carnegie Mellon, probably in late 1983, in the mid- 
dle of a World Chess Championship cycle. At the time I had no 
idea who the big names in chess were. Mike had to explain to me 
that Anatoly Karpov was the World Champion. Among the possi- 
ble challengers was Victor Korchnoi, who had lost two 
championship matches to Anatoly under, according to some, high- 
ly unfavorable conditions. Victor had defected and his family in 
the USSR suffered some persecution as a result. There was a rumor 
that Victor's family would be tortured if he won either of the 
matches. Mike thought Victor would be the most likely challenger, 
but he did not believe that he could win against Anatoly. Mike 
then mentioned the name Garry Kasparov, describing him as a 
flamboyant attacker with lots of talent. Mike thought that Garry 
would be a more exciting world champion than Anatoly, who was 
generally considered a little bit boring. But, again, Mike did not 
think Garry, being too young and too inexperienced, would be 
able to beat Karpov either. 

In 1984, Garry emerged from the Candidate matches 3 as the 
challenger to Anatoly. On 9 September 1984, the first World 
Championship match between them began. Whoever scored six 
wins first would be declared the winner of the match. Anatoly 
quickly scored four wins in the first nine games, while Garry was 
not able to win a single game. It looked bleak for Garry. Then he 
dug in, and drew game after game. Weeks grew into months, and 
the reporters started to leave. But the match went on, and Anatoly 
scored another win in game 27. Garry won the 32nd game and 
then the 47th and the 48th games. 

On 15 February 1985, the FIDE (the international chess feder- 
ation) president Florencio Campomanes announced "the 
termination of the match without a conclusive result". Garry com- 
plained that Campomanes was in league with Anatoly and was 
stopping the match to give Anatoly much needed rest. A new 24- 
game match, with the initial score of 0-0 and with draws counted 
as half a point each, would begin on 2 September 1985. In the case 
of a tie after 24 games, Anatoly would retain his title. I was busy 

3 The system for selecting the challenger for the World Chess Champion changed 
from time to time. For most of the recent history up to the mid 1990s, the chal- 
lenger was selected by a series of elimination matches also known as the 
Candidate matches. 



First Date with History 1 09 

working on the move generator chip and was not paying attention 
to the match. In this match Garry prevailed by the score of 13 
points to 11 points, and became the World Champion. 

By October 1989, Garry had successfully defended his title 
twice against Anatoly. He was officially about to become the first 
and, to the date of writing, the only chess player to have a FIDE 
rating of 2800 or above. None other than the legendary Bobby 
Fischer held the previous rating record of 2780. 

A few days before the scheduled match between Deep 
Thought and Garry, we got a call from CBS. "CBS This Morning" 
wanted to do an interview with Garry and the Deep Thought team. 
Jerry Present, who handled the public relations of the chess proj- 
ect for the Communications department at IBM Research, made 
the detailed arrangements with CBS. Murray and I stayed 
overnight at a hotel in New York City so that we could get to the 
studio on time. 

We woke early and arrived at the studio at about 5 o'clock in 
the morning, more than two hours before the show was to start. 
CBS wanted us to hook up Deep Thought remotely from the stu- 
dio and play a few moves during the segment. After setting up an 
IBM PC and testing the phone connection to IBM Research, we 
were led to a waiting room to have some breakfast. Garry had not 
arrived yet. Several monitors in the room were showing the on-air 
shows from CBS, other networks and the local TV stations. One of 
the CBS people stuck around and gave us some insider comments. 
This went on for a while. Then we heard the sound of footsteps. A 
familiar face appeared at the door. It was Garry. 

From the photos that I had seen, Garry appeared to be an 
intense person with a very high energy level. Seeing him in person 
confirmed the assessment. Photographs, however, could not really 
convey the intensity that came from his presence. Behind his dark 
eyes, there was a smoldering fire. Garry once referred to playing 
chess as "controlling chaos". Was there a fight going on between 
order and chaos behind those eyes? 

We introduced ourselves and chatted a little bit. A CBS person 
led Garry away to his makeup session. We followed shortly. After 
the makeup sessions, our conversation naturally turned to chess. 
Garry was about to overtake Bobby Fischer as the highest rated 
chess player ever, and we talked about the validity of comparing 
ratings from two different time periods. Kasparov declined to com- 
ment directly about the relative chess playing strength between 



110 Chapter 7 

himself and Bobby. We did not bother to talk about the likely out- 
come of the match between Garry and Deep Thought — there was 
never any doubt that he would win it easily. But what did Garry 
think his chance was against the machine that we were going to 
build at IBM? He made an interesting observation. 

When the team joined IBM, we were planning to build a chess 
machine that could search at least 100 million positions per sec- 
ond, preferably a billion positions per second. Garry thought that 
a machine searching 100 million positions per second would be an 
interesting opponent, and probably would beat the majority of the 
Grandmasters. But even for a one billion positions per second 
machine, Garry stated, "It might be able to beat the rest of the 
Grandmasters, but not Anatoly and me." He then explained that, 
in his opinion, he and Anatoly Karpov were the only real profes- 
sional chess players in the world, and that they were the only ones 
able to do proper preparations for the big match. Anatoly was rated 
2730, and the players right behind him were Timman, Gelfand and 
Ivanchuk. They were rated at about 2680, slightly behind Anatoly, 
but not by a full class (200 rating points). If one just looked at the 
rating scale, there was no reason to believe that the new machine 
would fall in the narrow range between Anatoly and the other top 
GMs. So what did Garry really mean? 

There are two plausible theories. 

One was that Garry knew subconsciously the machine would 
be a serious challenge but, refusing to bow to the inevitable, he 
placed Anatoly between the machine and himself as a buffer. This 
was Murray's theory when he first read about Garry's "Not Karpov 
and me" statement in Friedel's letter. Throughout the history of 
computer chess, chess players had made similar assertions but 
Deep Thought smashed the supposed barriers all the way up to 
"strong Grandmasters". It was interesting to note that, during the 
1997 match between Deep Blue and Garry, several Grandmasters 
claimed that Deep Blue was playing at around number X in the 
world. The number X happened to be slightly below the ranking of 
the specific Grandmaster making the claim. Some of the 
Grandmasters making the claims probably would have had their 
hands full with the old Deep Thought. 

The other plausible theory was that Garry really believed good 
preparation would make a huge difference in match performance. 
Actually, we believed this as well. In fact, the design speed for the 
"ultimate chess machine" was specified so that the machine would 



First Date with History 1 1 1 

play at about 200 points stronger than an unprepared Garry, or 
whoever the World Champion might be. We had good reasons to 
overshoot. A well-prepared IM David Levy was able to slaughter 
Cray Blitz, which was supposed to be comparable to David in chess 
strength, by 4 to 0. So it was true that good preparation could make 
a huge difference against a slowly changing computer opponent. 
But why "Karpov and me"? 

There was no fundamental reason why the other 
Grandmasters could not make comparable preparation but Garry 
and Anatoly did have more resources at their disposal than other 
Grandmasters. Grandmaster Joel Benjamin, who worked with us 
during the two Deep Blue matches against Garry, observed that 
Garry and Anatoly had lots of "slaves". Numerous Grandmasters 
wanted to be in the Champion's camp so that they could eke out 
a decent living or, in some cases, learn what it would take to be a 
World Champion. It was very difficult for a Grandmaster in the 
USSR to make a good living unless he was right at the top, and the 
collapse of the USSR and the subsequent loss of state support made 
the situation even more acute. Both Garry and Anatoly, each being 
a multimillionaire, became magnets to the ex-USSR Grandmasters. 
Western Grandmasters, like Jan Timman from the Netherlands, 
would have a much tougher time getting the same quality help. 
Furthermore, Garry and Anatoly had both successfully defended 
their titles. Who is to say that they don't have some secret formu- 
lae for top level preparation? 

The live TV segment went smoothly. Garry and Deep Thought 
played a few token moves. The rest of the interview, not surpris- 
ingly, was centered on Garry. Murray and I were, after all, just 
insignificant students fresh out of graduate school. My impression, 
from what he said, was that he saw himself as defending the 
human race against machines. When I looked at my fellow team 
members, 1 saw human faces. We were simply approaching the 
problem of playing chess from, shall we say, a non-traditional 
direction. Maybe Garry needed to pump himself up for the match- 
es to come. Well, that was fine with me. We wanted him to play at 
his best — okay, though perhaps not for the coming Deep Thought 
match where we were overmatched to begin with. 

Who Wants to Win? 

There was no doubt that Garry wanted to win the small exhibition 
match against Deep Thought with a perfect 2-0 score. We, on the 



112 Chapter 7 

other hand, were only concerned about not embarrassing our- 
selves. Losing by 0-2 would not be a big deal for us; to be honest, 
two losses by Deep Thought might even be the most desirable out- 
come. We certainly did not want our new employer to have the 
unrealistic expectation that we were ready to tackle the World 
Champion for real. Shelby Lyman, however, was not sure what 
outcome he wanted to see. A 2-0 win by Garry would be fine, but 
any half-point draw by Deep Thought would create huge head- 
lines. 

On Saturday 21 October 1989, the day before our first over- 
the-board encounter with Garry, we visited the match site and 
checked out the facilities. The match was to take place at the New 
York Academy of Art, which seemed a strange place for a comput- 
er chess match. When we walked into the playing room, we were 
greeted by an unusual sight. There were paintings and sculptures 
everywhere. In one corner was a giant hand, roughly the size of a 
love seat, possibly made from gypsum. Was it supposed to be the 
hand of fate? Or did it symbolize Deep Thought, the tool we had 
created with our hands, in its struggle with Garry? 

The playing room was quite cramped, especially after all the 
chairs for the spectators were added, giving an intimate feeling. 
Garry had a room behind the small raised stage to rest in and 
there was also a small area behind the stage that he could stroll 
around while he was thinking. A few reporters were floating 
around even though no press conference was scheduled for that 
day. We set up an IBM PC, hooked up the phone line, and tested 
the connection to Pittsburgh. One of the reporters wanted an 
interview. It turned out he was from Sports Illustrated. We were not 
in Kansas any more. 

After the interview with Sports Illustrated, we went upstairs to 
find Shelby Lyman. The commentary room, where the spectators 
could hear the live commentary, was on the second floor, but it 
was empty. When we found him, Shelby was quite agitated. It 
appeared that Garry had refused to talk to one of the reporters who 
appeared to have done something to upset him. Shelby and a few 
other people were trying to make Garry change his mind. 1 was 
curious as to whether Garry's mother, Clara Kasparova, was around 
but Shelby indicated that Clara had not come to New York. So, as 
I suspected, Garry did not consider the match a major event. 
Clara's husband had died when Garry was still a little boy and he 
and his mother had a very close relationship. In his autobiography, 



First Date with History 113 

Garry said, "The most important thing is that I can talk to her [his 
mother] as I can to nobody else." If it had been a major event, 
Clara would have shown up to cheer her son on. Garry, however, 
still did a thorough preparation for the match, as we would soon 
find out. 

When we arrived at the venue on Sunday, we saw something 
that you don't normally see for a chess match. The line of people 
waiting to buy tickets circled an entire city block. Around thirty to 
fifty reporters were milling around Garry. A few noticed us and 
started asking us questions. After a while, the reporters got seated 
and ready for the pre-match press conference. Murray and I had a 
brief discussion and agreed that if Deep Thought's assessment of 
the position dropped below -700, which was roughly 5.5 pawns 
down 4, then we would resign. After the discussion I looked up and 
saw Garry smiling broadly at us; he had overheard the conversa- 
tion. 

Besides the reporters, several of the usual suspects from com- 
puter chess events were also present. Monty Newborn, the 
perennial organizer of the ACM Computer Chess Championship, 
was in the front row. So was our former rival, Dr Hans Berliner, 
apparently not wanting to miss the historic spectacle even though 
he probably wished that Hitech was the computer competitor. 
Jonathan "Almost Winning" Schaeffer, author of the chess pro- 
gram Phoenix, was also present. (Deep Thought had played 
Phoenix three times, winning all three. But according to Jonathan, 
Phoenix was always "almost winning".) Mike Valvo, the tourna- 
ment director for the ACM and the World Computer Chess 
Championship, also showed up. 

Deep Thought got White in the first game, and I was the oper- 
ator. Murray manned the commentary room, relaying the moves 
from a PC. The PC, in turn, was following the game log on the 
computer in Pittsburgh. 

The game started well for Deep Thought. Then I noticed that 
something strange was happening. For several moves in a row, 
Deep Thought wanted to castle its king, but then switched to 
moves that did not make any sense. Eventually, Deep Thought 
decided to castle for real, but Garry already had an edge as the 
result of the strange moves. 

4 A pawn was worth 128 points on the Deep Thought scale. On the Deep Blue 
scale, a pawn was worth 100 points. 



114 Chapter 7 

Three weeks later, Deep Thought played Phoenix in the ACM 
Championship. The same behavior happened again, and Deep 
Thought was objectively lost against Phoenix. (So Jonathan was 
correct in his "almost winning" assertion this time.) Luckily for us, 
Phoenix was only "almost winning", and Deep Thought won the 
game after a long struggle in the endgame. Thomas found the 
problem after the wakeup call from Phoenix. In the rush to get the 
three-computer version working in time for the Garry Kasparov 
match, a subtle bug had been introduced. If the castling move was 
made immediately in the software portion 5 of the search, then no 
castling bonus was awarded to Deep Thought's evaluation of the 
position. But if the castling move was delayed, and played into the 
hardware part of search, then Deep Thought would get the bonus. 
So Deep Thought would go to extreme contortions to delay 
castling until it was in the hardware portion of the search. In the 
Phoenix game, Deep Thought never did get to castle as a result. 
Jonathan had a very pleasant time for most of the game at my 
expense. In his own words, I was "fidgeting, scowling, con- 
cerned" 6 . But after the position turned in Deep Thought's favor, I 
was "giddy, smiling and exuberant". Jonathan took the game quite 
hard and soon gave up on computer chess ~i. We stopped using the 
three-computer version of Deep Thought permanently after the 
Phoenix game. 

Anyway, in the first game against Garry, Deep Thought's 
pieces were seriously misplaced after the ridiculous contortions. 
Garry slowly increased his positional advantage and, after 52 
moves, when Deep Thought's evaluation dropped below -700, I 
resigned for Deep Thought. The game was lost way before that. 

While the first game was going on, a filming crew for the PBS 

5 In ChipTest, Deep Thought, Deep Thought II and Deep Blue, the positions most 
distant from the game position on the chessboard are searched by the chess 
hardware. These machines use special purpose chess hardware running under 
the control of a general purpose computer. For a 12-ply search, the first eight 
plies might be searched by the software on the general purpose computer, and 
the last four plies might be searched by the chess hardware directly. 

6 I tried to keep a poker face when operating the computer against human players 
but in games against computers, there was no need to hide my true feelings. 

7 Jonathan moved his attention to computer checkers, and eventually, with a few 
colleagues, created a World Championship level checkers program named 
Chinook. Chinook's story is described in Jonathan's excellent book One lump 
Ahead: Challenging Human Supremacy in Checkers. 



First Date with History 115 

TV show "Nova" was taking shots of Garry. After the first game, the 
Nova producer mentioned that Garry was staring at me "as if you 
were the enemy". I had been so intent on not making operator 
errors that I had not noticed — sometimes, ignorance is bliss. 

Garry, Murray and I traded places for the second game. 
Murray got to face Garry, while I relayed the moves to the com- 
mentators. Garry had White. 

Before the match, Garry had requested that we send him all 
of Deep Thought's public games. We gladly obliged. We weren't 
really taking the match seriously. If it had been a genuine match, 
we would have done more than just give Garry the public games. 
We would have gone into training, prepared new openings, and 
drastically changed the program or even the hardware, if need be. 
Human chess players do the same type of secret training when 
they go into World Championship matches. Usually, both camps 
shy away from serious public appearances for, say, six months 
before the big match. Obviously, neither camp would volunteer 
any surprises that they had prepared for their opponents. This was 
the case when Garry played for the World Championship in his 
matches against Anatoly Karpov, Nigel Short (Great Britain) and 
Vishwanathan Anand (India). In this match, however, Deep 
Thought was an open book to Garry, and we paid dearly for this 
openness, or rather, lack of secret training, in the second game. 

Garry used a move sequence that busted Deep Thought's 
opening for the Black side. The position right after the opening 
was difficult for Deep Thought, to say the least. Dr Berliner liked 
the position for White so much that he prepared and extended 
Garry's line for use by Hitech against Deep Thought. Three weeks 
after the Garry Kasparov game, Hitech played the White side of the 
same opening against Deep Thought. Deep Thought was lost right 
after the opening, but Hitech was not strong enough to finish off 
Deep Thought. Hitech was "almost winning". But once it was out 
of its opening book, Hitech immediately went astray. Deep 
Thought turned the tide and won the game. We did not have such 
luck against Garry. The castling bug, which existed in both games 
against Garry, did not help either. Deep Thought realized that it 
was in trouble right out of the opening, and it was straight down- 
hill from there. Deep Thought never thought it was ahead in the 
game even when it was up in material at one time. The on-site 
commentary, however, was asserting that Deep Thought believed 
it was ahead. One book that described the game even went as far 



116 Chapter 7 

as saying Deep Thought's evaluation was such and such ahead for 
it. Based on Deep Thought's highly negative evaluation of its posi- 
tions during the entire game, I believe that Garry simply busted 
Deep Thought's opening. Murray resigned for Deep Thought on 
move 37, but the result of the game was never in doubt after the 
disastrous opening. 

The thoroughness of Garry's opening preparation for the sec- 
ond game was an eye opener for us. Deep Thought never got to 
really play the game. It lost the game to Garry's home prepara- 
tion 8 . The easy win in the second game, however, might later have 
clouded Garry's judgment of what he could expect in the two 
matches he played against Deep Blue. 

Mr Karpov Will Have White 

Garry Kasparov's victory in the match against Deep Thought gave 
him by far the largest media coverage ever in the United States. 
After the match, Andrew Page 9 , Garry's personal manager, was 
extremely happy about how well Garry was received. Garry took a 
risk i° by agreeing to the heavy discounted appearance fee, but it 
would be an understatement to say that the publicity he received 
was well worth the discount. In certain aspects, Garry was getting 
even more coverage from the Deep Thought match than his World 
Championship matches. 

Garry's media bonanza did not go unnoticed by the other 
chess players, and Deep Thought would soon take on three other 
chess players in exhibition matches. The last match of the three 
was against German Grandmaster Helmut Pfleger. The match 
resulted in two draws. The other two matches had more exciting 
results. 

The first of these matches took place in December 1989, about 
two months after the Garry Kasparov match. The arrangements for 
this match, against International Master David Levy, were actually 
started before the Garry Kasparov match. David was best known 
for his taunting of computer chess researchers. In 1968, he had 
placed a bet with several computer, scientists that no computers 
would beat him in a match within the next 10 years. He collected 

8 Most likely it would still lose the game even if Garry had not busted its opening. 

9 Garry fired Andrew Page after the 1997 Deep Blue match over financial issues. 

10 Okay, $10,000 for one day's work is a risk that most people would undertake 
without any consideration. 



First Date with History 1 1 7 

the wager when Chess 4.5, a chess program from the Northwestern 
University, lost to him in 1978. Omni magazine then set up a 
$4000 Omni Prize, and David added another $1000 to the pile for 
the first computer to defeat him in a match. The Cray Blitz team 
tried to claim the Omni Prize in 1984, but was soundly defeated by 
David with the perfect score of 4 •. 

When David played Cray Blitz, he adopted what he called the 
"Do nothing, but do it well" strategy. He deliberately avoided giv- 
ing Cray Blitz targets, while keeping the board position closed so 
that the chances of immediate tactics were remote. The strategy 
worked exceedingly well against Cray Blitz. For that matter, it 
worked quite well against a lot of commercial chess programs, even 
at the hand of a weak chess player such as myself. Against David, 
Cray Blitz would muddle along until he set up some unstoppable 
long-term tactics and then Cray Blitz would crash and burn. Deep 
Thought had no problems spotting the long-term tactics used 
against Cray Blitz but, given that David was able to dispatch Cray 
Blitz with great ease, we were quite concerned before the match 
took place. When Deep Thought played Garry, it was acceptable 
that we did not score any point. Against David, we definitely want- 
ed to win. Between the Garry Kasparov match and the David Levy 
match, we added thirty new terms to Deep Thought's evaluation 
function in preparation. 

The match against David took place after Deep Thought's near 
disaster against Phoenix, and we were back to the single computer 
with two chess processors. The match was played in London. Peter 
Jansen, who became involved with the Deep Thought project in 
early 1989 u , was the operator. In theory, the match could have 
been played using the machine in IBM, but since only the 
machines in Pittsburgh were available to all the Deep Thought 
team members, we used a workstation at Carnegie Mellon for the 
match. 

Much to our relief, Deep Thought was able to do to David 
what he had done to Cray Blitz. David was gracious after his 0-4 
drubbing, and did not make excuses for losing. In his own words, 
"I was not feeling particularly unwell; I was not sitting too close to 

11 Peter's main contribution to the Deep Thought project was endgame databases 
for selected pawn endings. He did not get actively involved in the programming 
of Deep Thought. Peter was more a user of the chess machines, using ChipTest 
extensively in his PhD thesis research. It had been left behind at Carnegie 
Mellon when Murray, Thomas and 1 went to IBM. 



118 Chapter 7 

the heater; there was no disturbing noise outside the playing room; 
the lighting was not poor; the computer operator, Peter Jansen, 
was not smoking." 

Deep Thought's next match opponent was none other than 
the former World Chess Champion Anatoly Karpov. With Garry 
getting so much publicity as a result of the Deep Thought match, 
Anatoly probably wanted his share of the attention. Was this the 
reason why he agreed to the invitation from the Harvard Chess 
Club to play Deep Thought? I don't know. I do know that the invi- 
tation from Daniel Edelman, the president of the Harvard Chess 
Club, for Deep Thought to play Anatoly did not come until after 
the Garry Kasparov vs. Deep Thought match. 

Before his match Garry had requested the game scores of all 
of Deep Thought's public games. Anatoly did not make the same 
request and I assumed that he must have had the game scores 
already. I could not have been more wrong. 

International Master Mike Valvo had been doing some chess 
work with Grandmaster Ron Henley for over a year. Ron was one 
of Anatoly Karpov's seconds (chess advisors) and, being an invest- 
ment trader, he was also Anatoly's financial advisor. Mike Valvo, of 
course, was the tournament director for the ACM Computer Chess 
Championships, and thus was quite familiar with Deep Thought. 

The match against Anatoly was supposed to take place on 
Friday 2 February 1990 at Harvard University. At about 9 pm on 
the night before the match, Mike visited Ron to deliver a laptop 
computer containing ChessBase databases for Anatoly's upcoming 
match against the Dutch Grandmaster Jan Timman. Mike was sur- 
prised when Ron requested Deep Thought games for Anatoly to 
look at. It turned out that Anatoly had not prepared for the Deep 
Thought match at all, even though the match was very likely the 
most important event during his short trip to the United States. 

The first time I met Garry, he made the statement that only 
he and Anatoly were the true "professional" chess players. Well, 
compared to Garry, Anatoly was a quite different type of profes- 
sional player. Garry studied all of Deep Thought's public games 
and its weaknesses, including its opening. Anatoly did not do any- 
thing more than a superficial study before his match with Deep 
Thought. 

When the match was arranged, the Deep Thought team was 
surprised that it was going to be a one-game match. Chess match- 
es are usually played with an even number of games so that both 



First Date with History 1 1 9 

players get to play the same number of Whites and Blacks. The tra- 
dition for one-game matches was to select the playing color by a 
coin flip or some other chance event. The Harvard Chess Club 
organizers and ourselves all expected that the color would be cho- 
sen at random before the game, especially since it would be a 
one-game match. In hindsight, it seems that the Anatoly Karpov 
camp probably had a different idea early on. 

Below is an excerpt from an article written by IM Mike Valvo 
on the match: 

... Unlike both Fischer and Kasparov, however, Karpov is con- 
tent to play for a draw with Black and only press with 
White. 

It seemed to me that Karpov had made up his mind to play 
1. e4 after seeing the machine's handling of the Alekhine 
Defense 12 . I thought it was only fair to tell him that the 
machine could also play the Caro-Kann 13 , but I had no 
intention of saying anything more as I was privately root- 
ing for Deep Thought. I was taken aback at that point 
because Henley now phoned Danny Edelman and pro- 
ceeded to convince him that Karpov should be given the 
white pieces without a coin flip. Edelman was clearly 
resistant to the suggestion, but Henley was quite persua- 
sive. 

Murray and I were stunned when we heard the news. Having 
been a World Champion for many years, Anatoly was used to spe- 
cial treatment. Had he asked to play White at the time of the 
match invitation, I would have felt no qualms about it. It was pos- 
sible that we might not have agreed to play the match then, but it 
was also perfectly possible that we might. Anatoly's last minute 
demand left a very bad taste. Had we known that we would be 
playing the Black side earlier, we could have prepared a much bet- 
ter opening for Deep Thought. At high level chess, good opening 
preparation is crucial. Knowing well in advance which color you 

12 For the chess uninitiated, the Alekhine Defense is generally considered a some- 
what suspect opening for Black. Deep Thought had good results against weak 
players with the defense, but Anatoly Karpov had been known to mow down top 
grandmasters daring to play the Alekhine Defense against him. 

13 Caro Kann is a relatively sound defense for Black. Ironically, it was also Anatoly's 
favorite defense when he played Black at the time. 



120 Chapter 7 

will play against a specific opponent can be the difference between 
winning and losing. We were very unhappy about the turn of 
events, but we agreed to the demand in the end, for two reasons. 
First, we did not want to make a scene. The Harvard student group 
had put in a lot of work to make the event happen and we felt 
obliged to help them make it a success. Second, we were no longer 
only representing ourselves. Had it been entirely up to us, we 
might have said, "Good bye, nice meeting you, Mr. Karpov." 
Everything happened too fast, we had to make a snap judgment, 
and being forced to play Black seemed to be the least of our wor- 
ries. 

Anatoly looked quite different from what I expected. Most 
descriptions portrayed him as someone who looked relatively del- 
icate. When I saw him for the first time on 2 February 1990, he 
appeared quite healthy, although a little bit tired from his trip. 
Anatoly had remarried shortly before the match, and married life 
seemed to agree with him. 

After an opening speech by Anatoly, the game was under way. 
As Mike Valvo surmised, Anatoly played 1. e4, the king pawn open- 
ing. Between the Garry Kasparov and the Anatoly Karpov matches, 
Deep Thought had gone through the David Levy match and one 
computer chess tournament, during which Ken Thompson had 
provided us with his opening book for the Caro Kann. Mike Valvo 
was aware of this arrangement, and hence his deliberation about 
whether to tell Anatoly about the Caro Kann during their earlier 
meeting. As it happened, in the two intervening events none of 
our opponents played 1. e4, so there was no public game where 
Deep Thought played 1. ... c6, the Caro Kann, against 1. e4. 
Anatoly did not show any visible emotion when I made the Caro 
Kann defense on the chessboard. Unlike Garry, Anatoly's facial 
expression was not as easy to read. 

The game was played at a faster time control than in the 
match against Garry. The Anatoly Karpov game was played with 
one hour to each player, while the Garry Kasparov match was 
played with 90 minutes to each player. Surprisingly, by move 16 
Deep Thought had a chance to make a sacrifice that would have 
placed Anatoly under tremendous pressure. Had the game been 
played with a regular time control, then Deep Thought would have 
made the sacrifice. The 3-workstation version of the program that 
played Garry potentially could find the sacrifice within the game 
time, assuming that we had got it working for the match. It is 



First Date with History 121 

amazing how a minor difference could have changed the history 
dramatically. 

Since Deep Thought could not find the sacrifice over the 
board, it played the second best move and Anatoly obtained a 
slight edge. On the 27th move, Deep Thought played a move that 
at first glance looked quite ugly and the audience laughed. After 
the game, Anatoly commented that that ugly move was the only 
good move. I could not tell one way or another. All I noticed dur- 
ing the game was that the ugly move played was one hundredth of 
a pawn better than a more natural looking move 14. 

Unfortunately, Lady Luck did not stay with Deep Thought 
very long. The game entered into a rook ending, where each side 
had a king, a rook and a bunch of pawns. The rook ending was 
objectively drawn. The game remained drawn for many moves, 
and at one point Deep Thought had a chance for a draw by repeti- 
tion, but it spurned the draw as it erroneously assessed the position 
to be in its favor. Anatoly, being the virtuoso endgame player, 
knew better and he found some resources in the position that Deep 
Thought did not understand. Deep Thought ended up losing the 
game after 65 moves. Anatoly was down to his last minute when 
the game ended, while Deep Thought still had about twenty min- 
utes. The audience was a little bit surprised when I resigned for 
Deep Thought, given the limited time that Karpov had left. In my 
view, it was an exhibition game, Anatoly obviously knew what he 
was doing, and there was no sense wasting everybody's time to 
force him to play the game to the bitter end. 

After the game, Mike Valvo said to the audience that the game 
could be viewed as a moral victory for Deep Thought. I viewed it 
not so much as a hollow moral victory, but rather a good chess les- 
son given by one of the top chess players in the world. We needed 
to do some solid work about Deep Thought's or its successor's 
endgame play. The design of Deep Thought II was influenced quite 
substantially by the Karpov game, and there were very few games 
that had similar effects on the design of Deep Thought or its suc- 
cessors. When all was said and done, Anatoly Karpov's chess ability 
had to be respected. If letting Karpov play White was the price for 
the lesson, it was not so bad a deal. 



14 Chess programmers usually don't admit it in public, but this happens all the 
time. The "brilliant" moves are frequently only "better" than the losing moves 
by the slightest difference. 



IBM 



CHAPTER 8 



We Need a New Name 



A Cultural Shock 

The move from the familiar environment at Carnegie Mellon 
University to the IBM T. J. Watson Research Center in Westchester 
County, New York took me a while to get used to. 

A couple of months before we joined IBM Research, we got a 
call from Randy Moulic, who was to be our manager. Randy was 
excited about the project and wanted to visit us in Pittsburgh 
before we joined IBM. The meeting was mostly social. I was expect- 
ing someone relatively formal and much older than we were, but 
Randy turned out to be quite cordial, and only about ten years 
older. In fact, my first mental response was "our new manager has 
a baby face". Randy was quite young at heart and had a sunny out- 
look on life. Whoever selected him to be our manager did a good 
job; he was probably closer to our way of thinking than most IBM 
managers. The meeting went well, and I thought that the transi- 
tion to life in IBM would be relatively smooth. I underestimated 
the cultural shock that I was about to receive. 

Life as a computer science graduate student at Carnegie 
Mellon was comfortable. The stipend from the Department was 
adequate, and the cost of living in Pittsburgh was relatively low. 
Everyone in the Department knew you to some extent, and you 
knew everyone in the Department at least by name. Anyone could 
have a great deal of autonomy and freedom if that was what they 
wanted. If I needed any help, I knew with whom to talk, or at least 
someone who might be able to point me to the right person. The 
Department effectively operated 24 hours a day. Even in the mid- 
dle of the night, you could always find some fellow students or 
occasionally some faculty members about. 

125 



126 Chapter 8 

Compared to Pittsburgh, Westchester County had much high- 
er living cost — the food at the supermarket was about 10-20 
percent higher, and the rent for my new apartment was about 
three times more than in Pittsburgh. It was a good thing that my 
new salary was much higher than my old student stipend. The 
sticker shock took some time to get used to, but it was not as hard 
as the change in the work environment. 

The Immigration Course at Carnegie Mellon did a good job of 
easing new graduate students into the new environment at the 
Department. IBM Research did not have a comparable process to 
help newcomers. The first few months of my life at IBM were spent 
dealing with the seemingly never-ending paperwork, and trying to 
cope with the IBM computing environment, which was quite dif- 
ferent from the computing environment in typical universities l. 
Furthermore, I no longer knew what the other people around me 
were doing. Everyone seemed to be busy with his own project, and 
there was very little communication outside one's own group. 
Most of the people also appeared to work regular hours. The build- 
ing was usually almost empty by 7 o'clock in the evening. 

However, the most amusing new experience was all the dis- 
cussion about the name of the program. When I chose the name 
Deep Thought, it seemed to be appropriate for the machine that 
would defeat the World Chess Champion. I had no idea that some 
people might have problems with it. The first hint of potential 
trouble came at the very first computer chess tournament that 
Deep Thought played in, before we even started the employment 
talk with IBM. It was at the 1988 ACM Computer Chess 
Championship. Tony Marsland, a professor from the University of 
Alberta, organized a panel session alongside the Championship 
and I was one of the panelists. For some reasons, when introduc- 
ing me, Tony used the name "Deep Throat" at least three or four 
times although immediately correcting himself each time. "Deep 
Throat" was the code name of the informer for the Watergate 
investigative reporters from the Washington Post, and it was also 
the name of a movie that you would not take your mother to see. 

After the Garry Kasparov vs. Deep Thought match, com- 
plaints of the Deep Thought name poured in from both inside and 
outside IBM. It appeared that there were a lot of people who could 

i The amount of paperwork is much lower in IBM Research these days. The com- 
puting environment is also closer to what you see in the outside world. 



We Need a New Name 127 

not help thinking about "Deep Throat" when they heard the name 
Deep Thought. It got much worse. Murray Campbell's wife Gina 
was working at a Catholic college. The President of the college was 
an elderly nun, and during a conversation with Gina, the Sister 
kept on using the name "Deep Throat". Houston, we have a prob- 
lem. 

At Carnegie Mellon, we chose whatever name we liked. We 
could not do that at IBM. Given that IBM was a big corporation, 
we had to be careful not to infringe on someone else's trademark. 
It also seemed a good idea to get more IBM researchers involved in 
the naming process, so the Communications Department 
announced a contest for the name of the future IBM chess 
machine. Of the submitted entries, the name "Blue Chips" was my 
favorite, but a computer company was already using that name. A 
few other names also had the same problem. In the end, the name 
"Deep Blue" 2 was chosen, submitted by Peter Brown, that same 
classmate who was partially responsible for our being hired by 
IBM. For winning the contest, Peter was treated to a nice dinner 
with us, and a chance to play Deep Thought. He declined the sec- 
ond part of his award 3 . 

The Short-term View 

When we joined IBM In 1989, it was one of the most profitable 
companies in the world, with very good financial years until 1993 
or so. As a graduate student, I was used to operating with the min- 
imum amount of outlay. We had a nil budget for ChipTest, and a 
$5000 budget for Deep Thought 4 . You had better be frugal when 
you have a budget as tight as we did. To me, it was already a big 
surprise that IBM agreed to hire a full team immediately, consider- 
ing that most of the critical work would be in hardware during the 
early phases of the project. So when our manager Randy Moulic 



2 Deep Blue is a play on IBM's nickname Big Blue, which probably came from the 
blue color of its trademark. IBM does not own the trademark for "Big Blue" and 
never uses the term to describe itself. 

3 A few years later, Peter Brown, with the help of a strong master, played a casual 
game against a very early version of Deep Blue Jr. and lost. Peter wanted to be 
the first person to "play it". 

4 The commercial chess programmers, however, always referred to both ChipTest 
and Deep Thought as "supercomputers", implying that they only lost to multi- 
million-dollar machines. 



128 Chapter 8 

suggested that we hire a Chess Grandmaster as a consultant, my 
immediate gut reaction was "What for?". My feeling was that we 
would indeed need help from Grandmasters for the final push, but 
we were not really ready to work with them yet, since the new 
hardware was still no more than a glint in the eye. 

What I did not realize was that we could no longer afford to 
take only a long-term view. Back at Carnegie Mellon, Deep 
Thought was not an official project and the faculty did not have 
high expectations about how well we would do. Before we joined 
IBM, we had our own self-imposed milestones to meet, but we did 
not have to maintain a constant presence in the computer chess 
world. In theory, we could go back to the drawing board for years 
on end without showing up at any computer chess event. We did 
not have this luxury at IBM. Our arrival there was a high-profile 
one. Within two months, we were on the front page of the New 
York Times and the Wall Street Journal as a result of the exhibition 
match with Garry Kasparov. Such a high profile came with a price. 
We suddenly had to compete regularly in public and win. 
Providing short-term performance became an important concern 
for us even if it might be in conflict with the long-term progress of 
the project. It was most likely this concern that drove Randy to 
suggest hiring a Grandmaster. 

Grandmaster Maxim Dlugy was born in Russia, but he came 
to the United States when he was quite young and received most 
of his chess education in the US. To some extent, Maxim was a 
cross product of both the American and the Russian cultures. His 
name came up when we talked with some of our chess friends and 
we invited him to come up to IBM Research for a visit. He was 
articulate and personable, and after some internal discussions he 
was hired as a chess consultant. Murray, being the chess expert of 
the team, would be the main person working with him. 

When Maxim came on board, we had no clear responsibilities 
for him. In the first few weeks, he would come in, play a few quick 
games against Deep Thought, losing most of them, and then make 
a few comments. Later on, he would bring in some test positions, 
some from his recent tournament games, and see what Deep 
Thought came up with. Occasionally, much to his chagrin, Deep 
Thought would find that he had missed chances to win the games 
outright. The comments from Maxim were potentially very useful, 
but we were not in a position to make use of them without the new 
hardware at least partially designed. 



We Need a New Name 1 29 

Deep Thought had an opening book that was not up to 
Grandmaster standard, even though it was playing at Grandmaster 
level. After getting useful but not immediately usable work from 
Maxim, we asked him to work on a better opening book. In this 
new way of working, Maxim would first bring in an opening book 
fragment that he prepared at home. Deep Thought would check 
out all the positions for errors in the opening book, Murray would 
screen for the obvious errors, and then Maxim would check the 
flagged errors and come up with alternatives. 

The Dlugy opening book, however, was never fully checked. 
In 1991, when Bankers Trust was recruiting top chess players as 
potential equity or currency traders, Maxim left IBM and joined 
them 5 . We used his opening book in a Deep Thought vs. the 
German national team match 6 , with mixed results. Part of this 
opening book came straight out of the Encyclopedia of Chess 
Openings (ECO). The ECO is largely a collection of works from 
Grandmasters and some of the them were said to have deliberate- 
ly laid opening traps in the sections that they wrote. Among chess 
players, it is well known that the ECO cannot be fully trusted. In 
one of the games against the German players, Deep Thought 
walked right into an opening trap. ECO assessed the position as 
playable for Deep Thought, but actually it was a dead loss. Had-the 
position been examined, we would have noticed that Deep 
Thought assessed it as a full pawn down without any positional 
compensation, which is usually a good indication that the position 
is lost. 

The Dlugy experiment was largely unsuccessful. His opening 
book was relegated to a backup role when Deep Thought II became 
fully operational. By the time of Deep Blue, most of the Dlugy 
analysis was no longer active. However, we did learn something 
from the process — how not to work with a Grandmaster. To work 
with them properly, we needed to set clear goals and establish 
some means to measure our progress. 



5 This happened at about the time when Deep Thought II first became partially 
operational. See also the story on Deep Thought II and the Greener Pasture. 
According to my sources, Bankers Trust hired two other chess players in addition 
to Maxim Dlugy. 

6 The match was billed as a Deep Thought II match, but Deep Thought was the 
player. 



130 Chapter 8 

Deep Thought II and the Greener Pasture 

Deep Thought II, also known as Deep Blue Prototype, served sev- 
eral useful purposes during its career from 1991 to 1995. It was the 
strongest chess computer until Deep Blue supplanted it in 1996. 
The decision to build Deep Thought II, in retrospect, was probably 
a mistake. Because of Deep Thought II, Deep Blue was delayed by 
about two years, although part of this delay was due to unforeseen 
personnel changes. On the plus side, without Deep Thought II, Joe 
Hoane, who became critical to the success of Deep Blue, might not 
have joined the team. 

The initial decision to build Deep Thought II seemed to be a 
good one. Deep Thought was still about 200 rating points stronger 
than the nearest computer chess competitors, and would win three 
out of every four games against the best opponents. But the Swiss 
tournament format of the major computer chess events meant that 
its chance of winning an event was only about fifty percent, even 
though it was always the heavy favorite. When we were just grad- 
uate students, winning computer chess events was gratifying but 
otherwise unimportant. In a corporate environment, there was 
tremendous political pressure to keep on winning, even though 
the statistics were stacked against us. Officially, the reason for 
building Deep Thought II was to use it as a prototype for exploring 
parallel search algorithms. But the main reason was really to open 
up a large gap between our computer competitors and us, so that 
the project could go on for an extended period while I struggled 
through the design of the new chess chip. 

From my point of view, there was another reason to build 
Deep Thought II. I wanted to force Thomas Anantharaman to 
rewrite the program for the chess machine. 

Thomas and I went back a long way. He was the first person 
besides me to work on the Deep Thought series of chess machines. 
We were quite close, but by the time we joined IBM we had been 
in a long running debate on how to program the machine. Prior to 
his ChipTest days, Thomas had written a software-only chess pro- 
gram. In order to speed up his program, he violated many of the 
precepts of good programming and, in particular, his program was 
riddled with many "goto" statements, computer commands to 
branch to another place in the program. It was a classic "spaghet- 
ti" code — a program with convoluted, intricate control flow that 
could not be easily followed by normal human beings. 

It was no longer necessary for Thomas to write chess programs 



We Need a New Name 131 

in the "spaghetti" coding style when the ChipTest hardware start- 
ed working. But since he was doing it as a hobby, I could not really 
force him to make the change. Beggars cannot be choosers. The 
ChipTest software became sufficiently unwieldy even for Thomas, 
and when Deep Thought was being built, he initially agreed to 
rewrite the software in a manner that the rest of the team could 
understand. But partly because of my delay in delivering the Deep 
Thought hardware, and possibly partly because of Thomas' work 
load related to his thesis research, the Deep Thought software was 
modified from the ChipTest software, and retained all of its 
"spaghetti" code. 

After Deep Thought won the Second Fredkin Intermediate 
Prize, Thomas decided to drop his original thesis topic and turn his 
work on Deep Thought into a thesis, but the "spaghetti" remained. 
By the time we joined IBM, Deep Thought had about 100,000 lines 
of code, mostly in the "spaghetti" coding style, and even Thomas 
had a hard time keeping track of how the software worked. Deep 
Thought II, with many more chess processors (up to twenty-four) 
than Deep Thought, would require brand new software, and 
Thomas would finally have a good reason to change the program. 

Things did not work out as I hoped. 

When we started the design for Deep Thought II, Randy hired 
a contract engineer to work on the hardware design with me. He 
did a creditable job initially, freeing me to spend most of my time 
on the new chess chip instead of overseeing the design of the new 
machine. Before completing his work, the contract engineer left 
for a job outside IBM. Suddenly, I had to take over the whole 
design. This diversion took several months away from the chip 
work. 

There were more problems to come. 

When we were both graduate students at Carnegie Mellon, 
Thomas was working in the Speech group. For many years, he 
was running two programs, named "ex" and "Ipc", on the same 
machine on which I was doing my chip design work. I had 
always assumed that the programs were part of his speech 
research — in particular, the "Ipc" program sounded like a "Linear 
Predictive Coding" program, which is a well-established method 
used in speech recognition, and it made perfect sense for him to 
work on it. I only found out when I was about to leave Carnegie 
Mellon that neither of his programs was related to his speech 
work. The program "ex" turned out to be a program for the stock 



132 Chapter 8 

exchange 7 , and the program "lpc" was used to predict the 
movement of the stock market. Thomas had had an interest in 
the stock market for a long while. 

When Andreas Nowatzyk was approached by the Wall Street 
firm with the plan of forming a startup in Silicon Valley, he asked 
Thomas to work for the startup on a part-time basis. Andreas even- 
tually turned down the deal but Thomas was sufficiently interested 
that he continued his own negotiations with the firm. After some 
time, he reached a deal, believing that there was no conflict of 
interest with his IBM job. Unfortunately, IBM lawyers disagreed, 
because the deal involved the transfer of intellectual property. 
Thomas ended up leaving IBM for the greener pasture of Wall 
Street, with a new salary more than twice what he had earned. 

I was tinkering with the idea of creating my own version of 
the chess program before Thomas left. I had some new ideas for 
improving the search selectivity of the program, that is, making 
the program search much more deeply along "interesting" lines 
than Deep Thought ever had done. I wanted to see the ideas 
implemented 8 but, given my hardware design work, I was not real- 
ly serious about writing the chess program myself. I was just 
talking and hoping that Thomas would pick up the ideas and run 
with them as he had done frequently when we were back at 
Carnegie Mellon. With Thomas now gone, I was forced to write a 
new chess program for real, myself 9 . 

Arthur Joseph Hoane, a staff programmer at IBM Research 
whom we called Joe, had previously expressed his interest in the 
chess project. With Thomas gone, we had an extra head count 10 
and we needed someone to help program Deep Thought II. Joe was 
a graduate of the University of Illinois at Urbana-Champaign, and 
we knew him to be a self-motivated, excellent programmer. Deep 
Blue was not an ordinary IBM project; it was a project transplant- 

7 Don't ask me what the program did. Thomas never explained it in detail. 

8 After the Software Toolworks Championship, IM Mike Valvo challenged Deep 
Thought to a two-game correspondence match that he won easily. The corre- 
spondence games strongly suggested that it might be a good idea to have the 
program search a great deal deeper along "interesting" lines. 

9 Murray was a decent programmer, but at the time, only Thomas and 1 could pos- 
sibly write the new program as we were the only two with intimate knowledge 
of the hardware, the parallel search, and the selective search. 

io A head count is a hiring position. 



We Need a New Name 1 33 

ed from outside IBM, and with team members who were self-moti- 
vated. Before the team joined IBM, we explicitly told the company 
that we would not join unless we were given the chance to build 
the "ultimate" chess machine. Unlike typical IBM projects, the 
manager of the Deep Blue project was filling a support role, and 
not the project leader. As a matter of fact, there was no project 
leader. It was critical that members of the team be able to do things 
on their own. After some discussion and lots of paperwork, Joe was 
welcomed into the group. However, he was not yet ready to step 
into Thomas' shoes, so there was no break for me. I still had to pro- 
gram the new parallel search code and the new selective search 
code for Deep Thought II. 

At Carnegie Mellon, I wrote a simulator for the parallel search 
algorithms described in my thesis. With this simulator as the start- 
ing template for the new program, I had an easier time. However, 
it was a slow process. Writing software was never high on my list 
of interesting things to do. 

Deep Thought II had up to twenty-four chess processors, and 
the new chess program was designed to use all of them. The first 
version of the program did not implement any of the selective 
extensions used in Deep Thought, nor did it implement the search 
ideas that 1 had. It took several months to complete. The new pro- 
gram was about 5-8 times faster than Deep Thought. In his thesis, 
Thomas claimed that the search extensions used in Deep Thought 
produced only a small improvement in performance, which would 
not have been enough to compensate for the 5-8 times advantage 
in speed that Deep Thought II enjoyed. Yet, Deep Thought was on 
even footing with the first version of Deep Thought II in head-to- 
head competition. Maxim Dlugy had already left for Bankers Trust 
by then, but he still had some IBM contract hours to fill and he 
played a game against Deep Thought II. After the game, he opined 
that Deep Thought 1 1 was strong, but its play seemed to lack pizazz 
when compared with Deep Thought. In Maxim Dlugy's eyes, Deep 
Thought II played solid chess, but it did not seem to be capable of 
playing moves that surprised Maxim as Deep Thought had. I could 
not say conclusively that Thomas was wrong in his thesis, but the 
lab tests seemed to indicate there were reasons to doubt. Maybe the 
improvement was bigger than Thomas had concluded. 

The new selective search ideas took a few more months to 
implement. At first, Deep Thought was still holding its own with 
Deep Thought II. We had a few public events scheduled for Deep 



134 Chapter 8 

Thought II but since Deep Thought seemed as good as (or better 
than) Deep Thought II and was also the more stable program, we 
opted to let Deep Thought play in the events. The organizers, how- 
ever, had already publicly billed the events as featuring Deep 
Thought II. As recounted earlier, during one of these events, we 
discovered the need of further work on Maxim Dlugy's opening 
book. By then, his contract had expired. Murray eventually gave 
up on the book and wrote a program to automatically generate 
new opening books, using the top players' opening repertoires 
gleaned from the public game records. One of the opening books 
that he generated was based on Anatoly Karpov's opening reper- 
toire. This was accomplished by extracting the most common 
moves from all of Anatoly's public games n. This opening book 
was subsequently used in Deep Thought II's public debut. 

For a while, it appeared that all my software work was for 
naught. Deep Thought II was no better than Deep Thought, 
despite all the extra processing power. The hands-on experience, 
however, proved invaluable and I certainly got to appreciate what 
Thomas did more fully. I also had a better idea about how to design 
the hardware to ease the software work. But, more importantly, I 
got some new insights into how to control the search more effec- 
tively. It was with these new insights that the final touch was 
added. The result was dramatic. Deep Thought II was suddenly out- 
classing Deep Thought. Deep Thought was already a chess tactical 
monster and yet, in game after game, Deep Thought II found deep 
combinations that escaped Deep Thought by a few extra moves. 
We had a new, number one chess program. 

Deep Thought II played its first public event at the 1991 ACM 
Computer Chess Championship, held in Albuquerque, New 
Mexico from November 17-20. Instead of using the Dlugy opening 
book, the "Karpov opening book" was used. Deep Thought II won 
all its games; however, in one of its Black games, we had a scare, as 
the Karpov book allowed our opponent the option of getting a rep- 
etition draw right at the opening — Anatoly was well known to be 
willing to draw as Black. To me, the most impressive game played 
by Deep Thought II in this event was against Cray Blitz. For over 

1 1 Why did Murray choose to use Anatoly's opening repertoire instead of Garry's? 
Well, Anatoly played with a relatively narrow and safe opening repertoire while 
Garry's was very wide, rapidly changing, and sometimes risky. It would have 
taken a great deal of effort to ensure that a "Kasparov opening book" was safe to 
use. 



We Need a New Name 1 35 

ten moves (ten moves each by White and Black, or twenty plies 
total), Cray Blitz had no inkling that it was getting killed, while 
Deep Thought H was predicting the entire game continuation and 
assessing the game as completely won. Cray Blitz was one of the 
fastest chess programs around, and to outsearch it by this wide 
margin exceeded my wildest expectation. 

Joe Hoane took over the chess program after the Albuquerque 
event. At the time 1 handed it over, Deep Thought 11 had only 
about 5000 lines of code, or roughly one-twentieth the number of 
lines of code in Deep Thought. The Deep Thought 11 code handled 
ten times more chess processors, and it also had new and possibly 
far more powerful search extensions. I was proud of the work, but 
the software effort took about a year of my time which would oth- 
erwise have been spent on the chess chip design. 

A Trip to Denmark 

Over the years, after every big event that ChipTest, Deep Thought, 
or Deep Thought II played in, I would post in the USENET news- 
group rec.games.chess my personal account of the event 12 . The 
following report on our trip to Denmark was adapted from one of 
the posts. There were other important chess events in the career of 
Deep Thought II, including a match win over Grandmaster Judit 
Polgar, one of the top players in the world and possibly the best 
female player ever 13 . The chess event in Denmark, however, was 
important for many other reasons. Historically, it was the first time 
that a chess computer defeated a National Team in a match. From 
the Deep Blue team's perspective, it was also the first time that we 
played in a serious match of classical chess (chess played at tour- 
nament time control), and in that sense, it was a dry run for the 
Deep Blue matches that were to come. The event, especially the 
games against the wily Bent Larsen, provided many useful chess 
lessons for us. There is one more very important reason why I 
included this event in the book. Joe Hoane had taken over the pro- 

12 After the first Deep Blue match, I wrote a personal account but decided against 
posting it when the Communications Department at IBM Research persuaded 
me that it was probably better to publish it as an article in a serious magazine. 
As preparations for the second match were underway, the article never saw the 
light of the day. For the second Deep Blue match, 1 did not bother to write any 
post-match report. 

13 The match was played under action time control (30 minutes per player for the 
entire game) and should not be taken too seriously. 



136 Chapter 8 

gramming of Deep Thought II for just over a year when the event 
took place. By this time, he had made significant additions and 
improvements to the program, and Deep Thought II had more or 
less become his program. But it was only in this event that I really 
got to see Joe's dedication to the project for the first time. Joe was 
sick during part of the event, but he hung in there throughout, 
worked long hours, and made numerous improvements to the pro- 
gram. Murray, Joe, and I really solidified as a team during this trip 
to Denmark. 

I was able to get back to the chip design after the Albuquerque 
event, but found that the earlier design work was no longer useful. 
It was too long a break and I lost track of what I had been doing. 
Meanwhile, our manager, Randy Moulic, had something else on 
his mind. IBM was forming a joint venture with another company, 
and Randy was going to be involved. Chung-Jen (C J) Tan was 
assigned as our new manager in the spring of 1992 but since 
Murray, Joe and I were all in the project for our own reasons, and 
effectively autonomous, the change in the management chain had 
very little impact on our day-to-day routine. To me, the project was 
my life and certainly close to the same thing for Murray and Joe. 
We would see the project completed no matter who the manager 
was. As far as we were concerned, the manager was a support role 
and, even if we were to get the pointy-haired boss from the Dilbert 
comic strip, we would have gotten the job done w. CJ played a fine 
support role as the team was largely unaware of what he or the 
bureaucracy was doing. The team was well insulated from outside 
disturbances. 

While I was restarting my chip design work, Murray discov- 
ered the Internet Chess Server, and Deep Thought II played 
anonymously on the server for a while before it was discovered. 
Deep Thought II became fairly well debugged during this period, 
and it was good preparation for our next event. 

In late 1992, IBM Denmark talked with us about arranging 
chess exhibition matches, possibly in connection with IBM 
Sweden. At some point, there was talk about playing Swedish 



H Okay, I will admit that it would have been very hard if the pointy-haired boss 
had been our manager. In all likelihood, the pointy-haired boss would have pro- 
claimed himself our fearless project leader and ordered us to provide him free 
donuts at all the meaningless meetings he forced upon us. 



We Need a New Name 1 37 

Grandmasters as well, but the final arrangement was to play only 
in Copenhagen, Denmark, and only against Danish players. 

Two separate official matches were scheduled, a four-game 
match against GM Bent Larsen and a four-game match against the 
Danish national team, with the last game against Bent also count- 
ed as part of the team match. So, there were actually only seven 
official games. Five additional one-hour exhibition games were 
also scheduled during the weekdays. The Copenhagen Chess 
Union co-sponsored the matches together with IBM Denmark. For 
this event, we hand carried the printed circuit boards for the chess 
processors to Denmark and shipped the rest of Deep Thought II in 
a large crate. 

The main event against the Danish players was to start on 
Wednesday 24 February 1993. We arrived early on Friday 19 
February in order to take part in a public event the next day with 
Garry Kasparov. We had dinner with Kasparov and the organizers 
in the evening of our arrival. During dinner, we first learned how 
well prepared the Danes were. Jens Nielsen, the creator of a well 
known test set of chess positions for computers, had made a note- 
book for the Danish players, containing the Deep Thought games, 
relevant articles, pointers about computers' weaknesses, and so on. 
We were concerned, but Deep Thought II was in pieces, and there 
were prescheduled weekend events. There was not much we could 
do before Monday. 

Garry dominated the dinner conversation, but every once in 
a while he withdrew into himself and just listened. Maybe it was 
his cold. But it could also have been the case that something else 
was on his mind. A few days later, during one of our match games, 
we received a fax that Garry Kasparov and Nigel Short, his British 
Challenger, had decided to leave FIDE, the international chess fed- 
eration. They were forming a new chess organization, Professional 
Chess Association (PCA), and would play their World 
Championship match under the auspices of PCA. Nigel did not 
like the conditions FIDE was imposing on the World 
Championship match without his consent, and with Garry in a 
prolonged dispute with the then FIDE president Florencio 
Campomanes, Nigel found a willing partner. 

Garry was impressive on Saturday during the public event, 
despite having a cold. We could not put together the full Deep 
Thought II for the show, but we managed to get a version with 
only two chess processors up and running. Garry went over one of 



138 Chapter 8 

his recent games against GM Predrag Nikolic. In this game, Garry 
gave up major material for an attack, that in his judgment, was 
clearly winning. Deep Thought II, with reduced processing power, 
assessed the position as unfavorable for Garry initially, and gave 
some alternative defenses for the other side. Garry then proceeded 
to bust the new defenses given by Deep Thought II in real time on 
the stage. Deep Thought II clearly did not have a clue about what 
was going on in the game position. I was not precisely feeling 
gloom and doom, but it was clear that I would have a lot of work 
to do when I got back to the United States. The problems involved 
were more or less mine alone — the new chess chip would have to 
"understand" the positions in this game 1S . 

We started setting up the full machine at the match site on 
Monday 22 February. The machine room was not adequately 
cooled (there was no air conditioning) and one of the chess proces- 
sors overheated and stopped functioning after a few hours. The 
ventilation was improved by Tuesday and a test game was played 
without further mishap. 

The machine was referred to as Nordic Deep Blue by the 
organizers. This was the first time that the Deep Blue name was 
used, albeit obliquely. As far as I am concerned, Deep Blue did not 
exist until it was completed in 1996 for its first match against 
Garry. The machine that actually played in Denmark was Deep 
Thought II, although with a reduced number of chess processors. 
In Deep Thought II's public debut it played with twenty-four chess 
processors but at this time we were using only fourteen processors. 
Besides the one processor lost from the overheating, others were 
lost in a freak accident in our lab. The air conditioner in our lab 
iced up and stopped working. The ice then melted, dripped on the 
chess processors and caused some of them to malfunction. The 
peak speed of Deep Thought II was down to about 4-5 million 
chess positions per second for the Copenhagen events. 

Day 1 (Wednesday, 24 February) 

This was a disastrous day. It was the first time that we fully real- 
ized the seriousness of our own lack of preparation and how well 

is Today, Deep Blue Jr., which has only one-thirtieth the computation power of 
Deep Blue, has no problem playing the sacrifice itself in about a minute. So I did 
learn something useful from Garry that day. The chess chip needed to recognize 
that certain piece configurations, in conjunction with open lines against the 
enemy king, could be very powerful in a king hunt. 



We Need a New Name 1 39 

prepared was our opposition. Frantic work followed after the 
games. 

The first game was against Bent Larsen. This was the first offi- 
cial game of the match against him, and not part of the match 
against the Danish team. Bent adopted a simple strategy that worked 
surprisingly well. Bishops, operating on diagonals, are more power- 
ful when the chess position is wide open with many "open" 
diagonals uninhibited by pawns. If the chess position is blocked by 
pawns, the knights might become more powerful than the bishops, 
even though bishops are normally superior pieces. Bent traded off all 
the knights belonging to Deep Thought II and tried to keep the posi- 
tion blocked. Deep Thought II retained a pair of bishops, which is 
usually considered an advantage. The problem was that we had 
never programed Deep Thought II to understand that it should trade 
some pawns to increase the scope of the bishop pair. Bent's strategy 
would not have worked if we had forseen this as a potential problem 
beforehand. Our opening preparation also left something to be 
desired. Bent won the first match game, and we were in a deep hole. 

We played two short exhibition games after the official game. 
Again both ended in disaster. Our opponents were well prepared, 
and there seemed to be something wrong with Deep Thought II. 

Day 2 (Thursday, 25 February) 

The night before, a few modifications were made to the program. 
Joe put in an all-nighter. Murray also worked late. I was mostly a 
spectator. With the exception of the debut of Deep Thought II in 
Albuquerque, when I wrote the bulk of the chess program, my 
main work would have been completed usually by the time of a 
chess event. Overnight, Joe found some problems in the search 
extension code of Deep Thought II. Prior to our visit to Denmark, 
Joe put some debugging code into the program for testing, and this 
code was active during the games on the first day. The debugging 
code caused the program to behave quite differently and seeming- 
ly much weaker. Joe located the problems and corrected them. 

Three exhibition games were played during the day, and Deep 
Thought II behaved much better than it had the previous day. The 
schedule for our Copenhagen visit was exceedingly tight, and with 
the continuous exhibition games, we did not have much time to 
make serious preparation for the official matches. The Danish team 
was probably using the exhibition games to look for additional 
weaknesses in Deep Thought II. 



140 Chapter 8 

The second match game, and first game of the team match, 
was played in the evening against 1M Henrik Danielsen of the 
Danish national team. Henrik was surprised by several sharp 
moves from Deep Thought II. Joe's late night changes seemed to 
have made a big difference. After one particularly unpleasant move 
from Deep Thought II, Henrik spent over fifteen minutes thinking 
about his response. The machine correctly predicted his response 
and was able to use the time he spent on the position to come up 
with a killer move that pretty much ended the game. If Henrik had 
responded quickly, Deep Thought II would have missed the killer 
move. 

So after two days, Deep Thought II was 0-1 in the Larsen 
match, and 1-0 in the match against the Danish team. We slept 
easier that night. 

Day 3 (Friday, 26 February) 

It was a quiet day. Both match games against Bent Larsen and the 
Danish team, ended in draws. 

Deep Thought II had good chances to win the second match 
game against Bent, but could not pull it off. He was under pressure 
throughout the game, and the game progressed into an ending 
where both sides had a rook and a bishop and Deep Thought II had 
an extra pawn. The problem was that the bishops were not on 
squares of the same color. It is well known that endings where the 
two sides have a bishop on the opposite color squares are notori- 
ously difficult to win for the stronger side, since each bishop can 
only attack half of the chess board. Deep Thought II knew about 
this; unfortunately, it did not know that having only one addi- 
tional rook on both sides did not change the situation much. 

The second game of the day was against the Danish team 
member 1M Carsten Hoi. Carsten achieved a good opening posi- 
tion in part owing to an oversight in our opening preparation. 
Deep Thought II played strongly, given the weak position, and pre- 
sented Carsten with tricky choices. In one position, Deep Thought 
II appeared to be offering him a free pawn. Carsten thought for a 
very long time before declining the pawn offer and the game 
ended in a draw. Almost everyone in the audience expected 
Carsten to take the pawn and win the game. He explained after- 
wards that he did not see anything obviously wrong with taking 
the pawn, but his intuition told him that it might be a trap. Post- 
game analysis with Deep Thought II suggested that Carsten's 



We Need a New Name 141 

intuition was excellent. Had he taken the pawn, either his knight 
would have been trapped or Deep Thought II would have had a 
very strong attack. 

Day 4 (Saturday, 27 February) 

Bent Larsen had White again. Murray could not find a good anti- 
dote to Bent's tactics in the first game so the old standby, the 
Center Counter Defense, reared its head again. Of course, in 
Denmark, it is better known as the Scandinavian Defense, and 
Deep Thought II happened to play the "Danish variation". We did 
not play the defense to honor our host. We were desperate. Shortly 
after the start of the game, Deep Thought II was in big trouble. 
Probably the only chess-playing entity at the match site that did 
not "think" that Bent was winning was Deep Thought II. But even 
the machine assessed its position as highly unfavorable. Bent, 
however, did have one psychological problem. Going into the 
game, he was probably expecting another slow positional squeeze. 
The position over the board was very good for him, but he would 
have to execute the attack perfectly. If Bent were facing a human 
player, the human opponent probably would have cracked under 
the pressure. If he were facing another computer, there would have 
been a very good chance that the computer would have made 
some critical mistake and gone down in flames. Bent executed the 
attack well for a while, but Deep Thought II walked a fine line and 
defended resourcefully. Slowly, his advantage dissipated. When the 
endgame phase was reached, it was Deep Thought II that had the 
better position. There were some debates as to whether the final 
position was winnable, but since Deep Thought II had no idea 
about the proper way to proceed, we called a truce. 

The team match game against the third Danish National 
Team member, GM Lars Bo Hansen, was perhaps the best game 
played by Deep Thought II in Denmark. One commentator went 
so far as to say that Deep Thought II played like a Super 
Grandmaster. Lars had had previous successes against the opening 
line played by Deep Thought II, but said after the game that he 
might have to rethink the whole line. The game against Lars was 
the first time that we saw Deep Thought II outplay a Grandmaster 
in a regular tournament game by superior positional play. The final 
coup de grace was a tactical combination, but the game was finished 
before that. After this game, Deep Thought II clinched the match 
against the Danish National Team, leading by the score of 2.5 to 



142 Chapter 8 

0.5. We were still behind Bent by 2 to 1, and we had one last shot 
to tie the Larsen match. 

Day 5 (Sunday, 28 February) 

The anchor game for both the Larsen match and the Danish 
National Team match was played on this day. Bent got into trou- 
ble early, underestimating a critical move. Deep Thought H 
probably had a winning edge at move twenty-five, but then it 
decided to go into another ending with bishops of opposite color. 
This time, both sides had a queen as well. After the second game, 
we increased the penalty for having an additional rook for both 
sides, but did not do so for having a queen for both sides. Did we 
goof? Deep Thought II was not able to make any progress. After a 
short while, we started discussing in the machine room whether 
we should just offer Bent a draw. IM Bjarke Kristensen came into 
the machine room very excited and started to explain to us that 
the position was a win for Deep Thought II, if the queens were 
traded off! We were shocked to hear that. Murray was the chess 
expert of the team, and he did not know about the possibility. 
Bjarke told us that once the queens were traded off, the pure 
"bishop of opposite color" ending would be won. Deep Thought II 
could then create two widely separated "passed pawns" (pawns 
that can no longer be inhibited or blocked by opposing enemy 
pawns), which would win despite the bishops of opposite color. 
The problem we had, however, was that Deep Thought II was 
explicitly programmed not to trade off its queen in this situation. 
We played on a little while longer, and then offered Bent the draw. 
After the match, thoroughly disgusted with this experience, I went 
through books on chess endings and added knowledge about all 
the "easy" endgame rules, including detailed rules on the "bishop 
of opposite color" endings, to the new chess chip. 

The final tally for the Larsen match was Larsen 2.5 and Deep 
Thought II 1.5. The result was much better against the Danish 
National Team: Deep Thought II 3 and Danish team 1 . The results, 
looking back today, are no longer important, but the match expe- 
rience and the chess lessons we obtained from the trip were 
invaluable. 



i. 



CHAPTER 9 



Bringing Up the Baby 



Picking the CNp Vendor 

At Carnegie Melton, I never thought about the problem of choos- 
ing a chip vendor for the project. There was only one choice, 
MOSIS, the MOS Implementation Service. MOS1S provided low 
cost chips to the academic world by sharing the cost among mul- 
tiple projects. NSF (National Science Foundation) and DARPA 
provided the funding for university MOSIS projects, and usually 
students had their chips fabricated for free. 

To use the MOSIS service, I had to do the custom layout of the 
chip myself. Every single transistor for the 36,000-transistor chess 
move generator designed in 1985 was drawn by hand on a com- 
puter. I also hand routed every single wire on the chip. In 1986, 
when Teiji Nishizawa, my Japanese friend, visited Carnegie Mellon 
on his way to a conference, he was surprised that I managed to 
pack as many transistors as I did on a 3-micron CMOS chip. The 
commercial automatic layout tools at the time apparently did not 
do a very good job. I just had no idea how poor it was. 

IBM Research had an excellent set of internal IC (Integrated 
Circuit) design tools that were said to be better than the commer- 
cial tools. IBM Microelectronics Division also had the most 
advanced IC processes, and still does. The problem was that when 
we started at IBM, IBM Microelectronics was set up for volume IBM 
internal production only. It was not set up to provide a silicon 
foundry service. Today IBM Microelectronics provides one of the 
best, if not thebest service — if you want to build the fastest micro- 
processors using a foundry, you call IBM Microelectronics. In 1989, 
however, all the IBM researchers I met suggested that I find an out- 
side chip vendor instead. 

143 



144 Chapter 9 

Among the US-based commercial silicon foundries, LSI Logic 
and VLSI Technology were the two main candidates. With both 
commercial silicon foundries, I no longer had to do the layout 
myself. All that I needed to do was to design the circuit, and our 
chosen vendor would complete the physical layout using their 
own in-house automatic layout tools. My initial thought was "This 
is going to be easy." The circuit design for the move generator chip 
used in ChipTest and Deep Thought took only a month. Okay, I 
would have to add in the one-month time used to verify the chip 
as well. That was still only two months. I reasoned, "The new chip 
should take no more than a year or two." 

After some investigation, I was leaning towards VLSI 
Technology. In the early 1990s LSI Logic was concentrating more on 
the "gate array" business. In a gate array chip, all the transistors are 
prefabricated and the chip is customized by creating the metal inter- 
connection layers at the last steps. The gate array approach reduces 
the time from sending out the design to receiving the chip, but at 
the cost of circuit density and design flexibility. The main alternative 
is the "standard-cell" approach, which gives higher circuit density 
(by fabricating the transistors only where they are needed) than 
achievable with the gate array approach. The layout of the standard 
cell chips, however, could be automated in more or less the same 
way as the gate array chips. Furthermore, for chips that require sub- 
stantial on-chip memory, standard cell chips offer much better 
density, and improved performance. VLSI Technology had a more 
mature standard cell technology, but the clinching argument was 
their promise to lay out the custom circuit for detecting repeated 
chess positions, based on the design given in my PhD dissertation — 
for a tidy sum of money, of course. I would provide VLSI Technology 
with detailed transistor level schematics for the custom circuit. 

During the evaluation of the tools from VLSI Technology, I 
did a trial layout of one square of a simplified chess move genera- 
tor. The trial layout was for a 0.8-micron CMOS process, which 
should have been about fourteen times denser than the 3-micron 
CMOS process that I used for the old move generator done in 1985. 
The trial layout generated by the automatic tool was only about 
twice as dense as my hand layout done in the 3-micron process. I 
was about seven times better at packing the transistors than the 
automatic tool. That was the good news. The bad news was that 
the automatic tool that I had to use was about seven times worse 
at packing the transistors than I was. 



Bringing Up the Baby 145 

The poor packing density of automatic layout was the main 
reason for my decision to do a custom layout of the logic for 
detecting repeated chess positions. I thought the decision a neces- 
sary one, but it was a bad one. I would not make the same decision 
today. The custom design of the repetition detection logic inter- 
rupted the normal design flow, added new complexity to the 
design and delayed the chip design significantly. I spent over half 
a year just getting the transistor level logic correct for the repeti- 
tion detector. The repetition detector was also the indirect cause of 
a traffic accident in which 1 was involved. 

In the late spring of 1993, a few months after our Denmark 
trip, I delivered the final design data for the custom repetition 
detector to VLSI Technology's office in New Jersey. On my way 
back, a careless driver ran his car through a stop sign from a side 
road, only about twenty meters in front of me. It was a life-alter- 
ing experience, but I would not recommend anyone to try it. 
When the other car suddenly burst onto the road in front of me, I 
thought that I was going to die and it would be the end of the proj- 
ect l. I managed to swerve my rental car sufficiently to avoid 
hitting the other car straight on. Had he seen me, and stopped in 
time, an accident would have been avoided. As it was, he went on 
without any regard to the road condition and a glancing collision 
took place. I was alive, but with the face bruised and cut by my 
glasses that were shattered in the impact with the air bag. I also 
had lacerations on my face and left arm. The occupants in the 
other car were uninjured. At a nearby hospital it was found that 
the cornea of my left eye had a nasty cut, and I had to wear an eye 
patch for a week. I did not really get back to serious work for about 
a month because of the problem with my eye 2 . By the time I got 
back to work, as a result of the prolonged break and possibly the 
trauma from the accident, I had a hard time understanding the 
existing design of the evaluation hardware. After spending about a 

1 When Andreas Nowatzyk read an early draft of this book, he was surprised at 
what I was thinking right before the accident. To me, the project was my life dur- 
ing this period. Also, 1 was returning from a business trip related to the project, 
and the project was on my mind. At that point in time, the survival of the proj- 
ect relied entirely on whether the new chess chip worked. Since I was the sole 
designer of the chip, if 1 had not survived the accident, the project would cease 
to exist. 

2 For over a year after the accident, I continued to have problems with my eye- 
sight from time to time. 



146 Chapter 9 

week going over the earlier design, I gave up and decided to redo 
the evaluation hardware from scratch. I had already spent about 
half a year on the evaluation hardware which was all wasted. 

It got worse. In late 1994, when discussions about a potential 
match between Deep Blue and Garry Kasparov were initiated, the 
first version of the Deep Blue chess chip was still being designed. 
There was still no sign of the custom repetition detector from VLSI 
Technology after over a year. With time getting short, the decision 
was made to drop the custom repetition detector from the design. 
So the six months spent on designing the detector went down the 
drain as well. In total, the one bad decision to create a custom rep- 
etition detector delayed the project by one year. A pricey decision, 
not just in terms of the wasted time, but also in terms of a vastly 
weakened chess machine. While detecting repetition does not 
seem very important at first sight, as you will see later, it can have 
a very subtle influence on a chess program's positional play. 

At about this time VLSI Technology decided to move the work 
to their Boston office. A few months later, when I delivered the 
first version of the complete netlist to VLSl's Boston office, sans the 
repetition detection logic, I saw the partial layout of the custom 
repetition detector. The layout was about five times bigger than I 
expected, with over 40 percent of the area occupied by two giant 
wire bundles crossing each other from one side to the other. An 
automatic tool would have caught the fact that the two halves 
could be interleaved like a comb, and the crossing wires should 
occupy very little area. So the custom layout somehow became 
much worse than an automatic layout. The engineer at VLSl's New 
Jersey office had been told that the wires should be interleaved, but 
somehow the information was never relayed to the layout techni- 
cian. We never would have accepted the custom layout as it was, 
even if we had the time to wait for its completion. 

Today, if I design another chess chip or any other chip, I 
would approach the chip vendor decision quite differently. The 
semiconductor industry has changed quite a bit since the late 
1980s. Then, chip designers designed the circuit with schematics, 
often using chip vendor specific tools. From the early 1990s, chip 
designers turned to language-based logic synthesis tools, describ- 
ing the chip in hardware description languages and relying on 
automatic logic synthesis tools to create the netlist. The chip could 
be designed without a specific chip vendor in mind. The designer 
could pick the chip vendor very late in the design process since 



Bringing Up the Baby 1 47 

there is only a small penalty to switch the chip vendor. Alas, when 
the chess project was started, I did not have this luxury. By the 
time synthesis tools became widely used, IBM was in serious finan- 
cial trouble and there was no money to acquire the new tools. The 
Deep Blue chess chips were designed almost entirely using circuit 
schematics even though there were far better design tools available 
near the end of the project. We also did not have many choices on 
the chip technology front as a result of using vendor-specific tools. 
The 1997 rematch between Garry Kasparov and Deep Blue was 
played using 0.6-micron CMOS chips, even though the industry 
had already moved on to 0.35 micron CMOS technology and 
beyond 3 . Anyway, the reasons for choosing VLSI Technology over 
LSI Logic would no longer be valid today. In particular, I would 
also avoid any custom layout like the plague, unless I have full 
control over it. 

At the time that we made the decision, choosing VLSI 
Technology was probably as good as any other choice. In the final 
count, VLSI delivered their end of the new agreement. So every- 
thing did turn out all right. There were better choices on the design 
tools later on, but we were in no position to make the choice, 
given the tight IBM budget at the time. By the time IBM's finances 
were back on track, the design was already close to completion and 
it was too late to change the tools. 

Murphy's Law 

Serious negotiations between IBM, Garry Kasparov, and the ACM 
(Association for Computing Machinery), which was the main sanc- 
tioning organization for the big match, were taking place in late 
1994. The match was to be held in February 1996 to coincide with 
the celebration marking the 50th anniversary of ACM. The design 
of the Deep Blue chess chip was near completion, but there was 
still the huge task of verifying its correctness. Murray and Joe 
agreed to drop what they were doing and help with the chip veri- 
fication. Deep Thought II was turned off and disconnected because 
we needed its host computer, which was our only machine with 
sufficient main memory to run the chip simulation. We were 
expecting to turn Deep Thought II back on for a few test games 

1 The chess chips would have been about twice as fast if we had had access to the 
better technology. The die area should reduce by about a factor of four, which 
would improve the chip yield quite dramatically. 



148 Chapter 9 

when the new chess machine became operational; otherwise, as far 
as we were concerned, Deep Thought II was no more. 

The ICCA (International Computer Chess Association) held a 
World Computer Chess Championship tournament triennially. 
They approached IBM to sponsor the event, to be held in Hong 
Kong in May 1995. We had won the event in 1989, but did not par- 
ticipate in 1992. If it had been entirely up to me, we probably 
would have skipped the event again. Murray, Joe and I were 
already swamped with the chip work as it seemed far more impor- 
tant to make sure that we would be ready for Garry Kasparov. 

However, our manager, C J Tan, had a different idea. Seeing 
the championship as a chance to add publicity to the match with 
Garry, he arranged to get the funding for the championship and 
agreed to participate. There was a slim chance that the new chess 
chip might be available for the Hong Kong event, but it soon 
became clear that it would not even be ready for fabrication by 
then. The only other choice was to compete with Deep Thought II. 

Jerry Present, our communications person at the time, want- 
ed to use the name Deep Blue Prototype instead of Deep Thought 
II. I wanted to use the name Deep Thought II. My reasoning was as 
follows. Deep Blue, the new machine, would be at least 100 times 
faster than Deep Thought II in effective search speed, and further- 
more, Deep Blue would have a far better grasp of positional 
concepts in chess. Comparing Deep Blue with Deep Thought II 
would be like comparing the sun with the moon. Well, I exagger- 
ate a little bit, but the difference in computation power was 
roughly of the order of a thousand to one, taking into account the 
far more complicated chess evaluation computation carried out on 
the Deep Blue chess chips. I was proud of what had been done, and 
did not want anything to be linked to Deep Blue unless it did use 
the new chips. Deep Thought II was still using the chess chips that 

I designed back in 1985. As far as I was concerned, Deep Thought 

II was a dinosaur about to become extinct. It was not going to 
usurp the name of our new megaton solar blaster, even if only par- 
tially. 

Jerry, however, persuaded C J to stick with the name Deep 
Blue Prototype. The Communications Department would like to be 
able to say that Deep Blue was the successor to the reigning World 
Computer Chess Champion, Deep Blue Prototype, at the time of 
the match with Garry Kasparov. In the "unlikely" case that Deep 
Thought II did not win the championship, then it was not really 



Bringing Up the Baby 1 49 

Deep Blue playing anyway. The problem was that, given the Swiss 
tournament format of the championship, the "unlikely" was high- 
ly probable. Despite the fact that Deep Thought II beat our closest 
competitors by about a three to one margin (or 200 rating points 
difference in computer vs. computer play) in lab tests, our chance 
of winning the championship was only about 50-50. There were 
twenty-four competing teams, and only five rounds of games 
would be played. With so few rounds, luck played a very important 
part in winning the tournament. We were the strongest, but there 
were about five or six teams just one class below, and another three 
to five teams only slightly weaker than they. The team, Murray, 
Joe, and I, all knew about our real chances given the format. Randy 
Moulic, our first manager, also would have known about the odds, 
having gone through a 1990 ACM Championship that we did not 
win outright. But C J had only been through the 1994 ACM 
Championship with us. In that one event, Deep Thought II won 
the title outright, even though we forfeited one game when a sud- 
den storm in New York knocked out the electric power to the lab 4 . 

Deep Thought II remained turned off until about a month 
before the Hong Kong event. Murray spent a small amount of time 
preparing it for the championship. The final verification work on 
the new chip was still going on, and we did not have much man- 
power to spare. We had to free up some of Murray's time, as he was 
the chess expert. We had never been so cavalier in our preparation 
for a computer chess event, save for ChipTest's first outing. Then, 
we did not care whether we won or not, and for this last outing of 
Deep Thought II, our main concern was keeping the date with 
Garry, nine months away. 

It seems that things usually go wrong when you are least pre- 
pared. As Murphy's Law predicts, everything that could go wrong 
soon went wrong. 

Murray and I, traveling with C J, flew to Hong Kong to be the 
on-site operators. Joe stayed behind to watch over Deep Thought 
II. It was the first time since the start of the chess project in 1985 
that I had been back to Asia, it was also my first trip to Hong Kong. 

The first hint of potential trouble exhibited itself when 
Murray and I went to the playing site, the Chinese University of 
Hong Kong. The Internet connection to the United States was not 

4 This event will be described in the next chapter. The first discussions I had about 
the first Deep Blue match took place at the end of the event. 



150 Chapter 9 

quite working; for that matter, neither was the connection to 
Europe. The three main universities in Hong Kong used a common 
carrier, which had a limited capacity, for all their Internet connec- 
tions. Before the event, the teams with computers on remote sites 
had tried out Internet connections into Hong Kong from their 
home countries, and they appeared fine. But getting out of Hong 
Kong using an Internet connection was almost impossible, at least 
during the tournament. It looked as though the remote teams 
might have to pay very expensive long distance phone bills. 

As it happened, the day before the championship IBM Hong 
Kong had announced an IBM Internet service for Hong Kong. This 
service was usable although the connection seemed to drop from 
time to time. It might have been a problem with the local phone 
service, since we were having problems when we dialed long dis- 
tance as well. (In 1997, 1 was back in Hong Kong doing a Deep Blue 
demo via the IBM Internet connection, and it went smoothly. So 
the communication problem we had in 1995 might not be with 
the Internet connection at all.) The phone line problem became an 
important factor in the last game played by Deep Thought II. 

Our first round opponent was Star Socrates, a parallel chess 
program from MIT, running on a multi-million-dollar supercom- 
puter — a machine about a hundred times more expensive than the 
workstation that Deep Thought II ran on. Star Socrates was con- 
sidered our most dangerous opponent. Its search speed was at least 
comparable to Deep Thought II. When Deep Thought II played 
Star Socrates the year before, I saw for the first time a program that 
actually "out-searched" Deep Thought II — it was reporting search 
depths larger than Deep Thought II's. Yet Deep Thought II had out- 
classed it easily. Would one year make a difference? Deep Thought 
II had been off line for over half a year, and the MIT folk were prob- 
ably not sitting idle. In 1994, Deep Thought II had played a regular 
chess opening, but with only one month of opening preparation 
for the Hong Kong event, Murray opted to play an irregular open- 
ing to take both programs out of their opening books. The game 
became a contest of raw chess playing strength. I was not too 
happy with Murray's opening choice; we had White, and yet we 
had a slightly worse position out of the opening. Deep Thought II, 
however, soon overpowered Star Socrates. 

For the next two rounds, Deep Thought II continued to play 
irregular openings; we won both games and led the field by one 
full game. I continued to mumble to Murray about the bad open- 



Bringing Up the Baby 151 

ings. In retrospect, I should have kept quiet. Murray chose to play 
irregular openings out of necessity — there was no time to do prop- 
er opening preparation. Deep Thought II was a better player than 
the rest of the field, and by taking other programs out of their 
opening books, Murray gave our opponents a longer rope with 
which to hang themselves. 

Our fourth opponent was WChess, one of our tougher oppo- 
nents. Deep Thought II played a normal opening this time. The 
position was even when both programs came out of their opening 
books. Deep Thought II soon found a way to win a pawn and had 
a winning position, but WChess had some counter play. Deep 
Thought II made one bad move and its position deteriorated. 
Almost immediately after this, Deep Thought II found that 
WChess had a way to draw the game. WChess did not see the draw 
at this point, but it played the right moves. Deep Thought II sud- 
denly found its bishop was close to being trapped and it had to 
give up material to keep the bishop. WChess was about to go ahead 
in material. It was a fast turn of events. At this point WChess saw 
the draw, and soon also saw that it would be ahead. Luckily for us, 
Deep Thought II had enough pull before its bad move, and despite 
the bad move we were able to hold the draw. After this game, I was 
a little bit uneasy about our decision to drop the repetition detec- 
tor from the design of the new chess chip. If Deep Thought II had 
had the ability to detect repetition in hardware, it would have 
found where the bad move would have led and played a better 
move instead. 

Nonetheless, we still led the field by half a game, and our last 
round opponent was Fritz, an easy opponent for us. Fritz's strong 
point was its tactics, but Deep Thought II was a far better tacti- 
cian; in fact, it won about nine out of ten games against Fritz in 
our pre-tournament tests back at IBM. I was exhausted on the way 
back to our hotel. The previous few days, I had been complaining 
about the bad openings. By this time, especially after the not so 
good result in round 4, I was of the opinion that the irregular 
opening approach had been a very good idea. Given the relative- 
ly weak positional sense of Fritz, an irregular opening should work 
really well. For once, I kept quiet when I should have spoken up. 
When I went to sleep that night, 1 was expecting Murray to cook 
up some irregular opening for Fritz. The next day brought a big 
surprise. 

Murray went to the playing site early and chatted with some 



152 Chapter 9 

of the other participants. One told him that Fritz had a horrible 
opening book, and had lost one of its games right out of the open- 
ing book. Murray took the comment seriously and decided to use 
a normal opening against Fritz. After all, if Fritz had been a piece 
of cake when we played irregular openings, it would be even easi- 
er if we played a real opening, especially given that Fritz had a bad 
opening book. Murray did not tell me about his decision before the 
game. Deep Thought II opened with a king pawn opening, and I 
gave Murray a quizzical glance. Fritz replied with a sharp line in 
the Sicilian Defense. A clever "transposition" (deliberate swapping 
of opening moves from their normal playing sequence) on the part 
of the author of the Fritz opening book took Deep Thought II, out 
of its own opening book. There were four main move choices for 
Deep Thought II, and two of them were castling moves that nor- 
mally take the king out of the center into the safer wing positions. 
Both castling moves, probably by design of the author of the Fritz 
opening book, were bad in this position. Deep Thought II assessed 
the position as favorable for itself and wanted to castle. Murray's 
face turned white. The position looked precarious for us. It was the 
first move out of the opening book for Deep Thought II, and it was 
spending extra time on the move. Perhaps there was some hope 
that Deep Thought II would play a safer move. Suddenly we got 
disconnected. According to Joe, who was watching the game from 
our lab in Hawthorne, Deep Thought II did switch to an alterna- 
tive move. But the new move never showed up on our screen in 
Hong Kong before the line drop, and we did not know about it 
until after the game. 

The MIT Star Socrates team was using a program that allowed 
them to reconnect to their chess program without the need to 
restart. Had we used the same program, we would have reconnect- 
ed, found the new move already played by Deep Thought II on its 
internal board, and played it on the real chessboard. As it was, we 
had to restart Deep Thought II, which now had less time to spend 
on the position. Compounded by the fact that it was started "cold" 
without the help of previous calculations and thus probably a fac- 
tor of two slower effectively, Deep Thought II could no longer find 
the more reasonable alternative move. 

After Fritz's next move, the evaluation of Deep Thought II 
dropped to a repetition draw. So it was a repetition detector prob- 
lem as well! If Deep Thought II had had a hardware repetition 
detector, it would have avoided the bad castling move, with or 



Bringing Up the Baby 153 

without the line drop. I was feeling even worse about the decision 
to drop the hardware repetition detector from the new chip. 

Our position against Fritz was getting bad, but it was not 
hopeless. We continued to have problems with the communica- 
tion line. We muddled along, losing precious time on our clock, 
but we were not busted yet. Then it hit. Deep Thought II found 
the move that it wanted to play, was losing, and entered into the 
"panic time" state where it would spend extra time to find an 
alternative move. Grandmaster Robert Byrne, the New York Times 
chess columnist who was the honored guest of ICCA, dropped by 
and whispered that Deep Thought II had to play a certain move 
to keep the position alive. The move intended by Deep Thought 
II was not the move that Robert suggested. We watched helpless- 
ly as Deep Thought II spun its wheels, trying to find a good 
alternative. The search appeared to be exploding — Deep Thought 
II was suddenly searching a much larger tree. Several of the alter- 
native moves were rejected. Would Robert's move be found in 
time? The clock was ticking. Time was up. Deep Thought II played 
its originally intended move as nothing better had come up. The 
game was effectively over. Fritz did not play the best attacking 
moves, but it was good enough to win the busted position, and we 
resigned for Deep Thought II. Fritz ended up winning the tourna- 
ment after beating Star Socrates, when it self-destructed in their 
playoff game. 

The tournament taught us several things. From my point of 
view, the most important was that if I ever got another chance to 
make a newer version of the chess chip, it must have a hardware 
repetition detector. There were other lessons too. We had to take 
opening preparations far more seriously, and we must have good 
backup plans whenever we played remotely. We should be able to 
reconnect to the chess program in the same way as the MIT team, 
and we should test out the communication lines thoroughly 
beforehand. We could not do anything for the hardware repetition 
detector problem before the match with Garry, but we certainly 
could do something about the communication issues. 

The decision to use the name "Deep Blue Prototype" also 
came back to haunt us. Several computer chess vendors renamed 
"Deep Blue Prototype" to "Deep Blue" when describing their 
result from the Hong Kong event. One went as far as saying that 
their program beat both Garry Kasparov (in blitz games) and Deep 
Blue (it never played Deep Blue). Although, to be fair to them, the 



154 Chapter 9 

vendors probably would have done the same even if we had used 
the name Deep Thought II. 

Four Hours to Spare 

Although we dropped the plan to use the new chess chip for the 
Hong Kong tournament, there was another event where it would 
have to be used. IBM Research was planning to open a new 
research lab in Beijing, China, and we were invited to play a friend- 
ly match with Xie Jun, the Women's World Chess Champion, as a 
part of the opening celebration. The Beijing match was scheduled 
for late September 1995. A clause in the match contract with Garry 
Kasparov stated that IBM could terminate the match contract on 
or before 31 October 1995. The team was not told explicitly that 
we had to do well in the Beijing match but, given the time prox- 
imity, the link was clear. 

In late August 1995, VLSI Technology gave us some disturbing 
news. None of the chips in the first batch passed the manufactur- 
ing test. They sent us some of the better behaving chips in early 
September. The Xie Jun match was scheduled for September 21, 
less than three weeks away. It was a very tight schedule even if the 
chips had passed. 

At the time we got the news from VLSI Technology, Joe had 
started porting the Deep Thought II software to the new hardware. 
I had received the printed circuit boards for the new chips a few 
weeks earlier and had done the basic tests on them. Now, Joe and 
I were in for some fun time together. 

The basic tests for the printed circuit boards simply checked 
whether the boards could be initialized properly. I had prepared 
more advanced tests for testing the chip with the board. None of 
the chips passed the simplest of the advanced tests that checked 
the connectivity of the bus wires to and from the chips. The prob- 
lem could be with the board or it could be with the chips. The 
basic tests for the boards used only 8 bits of the 32-bit 
MicroChannel bus in the IBM RS/6000 workstation. The advanced 
test used all 32 bits to check out the chip operations. It was possi- 
ble that there were something wrong with the 24 untested bus bits. 
Using an ohmmeter 5 , I checked the connections between the 
chips, and the MicroChannel bus on the board, and found that 
physically the wires were connected, but somehow they were con- 

5 An ohmmeter is an electric meter used to measure electrical resistance. 



Bringing Up the Baby 155 

nected incorrectly. It turned out that the RS/6000 workstation 
interpreted the 32 bits differently from the printed circuit board. In 
technical terms, this is the so-called "byte swapping" problem. The 
32 bits can be divided into 4 bytes, with each byte having 8 bits. 
The workstation assumed that the 4 bytes were ordered in a certain 
way, while the board assumed that the 4 bytes were ordered in a 
different way. Okay, this was not pleasant, but the problem could 
be bypassed by reinterpreting the 32 bits in software. 

After the software change, the chips started to pass some of 
my tests, but they all failed on one of the simpler tests. Worse, even 
the same chip failed differently from one test run to another. So, 
the problem was probably not with VLSI Technology's manufac- 
turing test setup. Something peculiar was going on inside the chip. 
I tried out all sorts of hypotheses. Since the chip worked during the 
pre-fabrication simulation, the most likely cause had to be some 
sort of electrical problem. The question was what type of electrical 
problem. In one of the trial tests, there seemed to be some pattern 
sensitivity in the chip's behavior. The chip appeared to behave dif- 
ferently depending on what was last read out from it. To test out 
whether there was indeed a pattern-specific sensitivity, I tried set- 
ting the last readout to be a 32-bit zero before doing a failed read 
operation from the chip. The chip now passed the test that it had 
previously failed. I then added a read operation that appeared to 
read out zeros reliably before every normal read operation, and the 
chip passed all the modified tests. I informed VLSI Technology 
about the finding and sent them the revised manufacturing test 
vectors, so that we could get factory prescreening on all the incom- 
ing chips as soon as possible. 

There was still the problem of what was causing the chips' 
strange behavior. I tried setting the extra read operation to read out 
32 bits of Is, and again the chip worked. I now strongly suspected 
that the cause of the problem was some sort of "cross coupling" 
among the 32-bit output bus wires on the chip. When one of the 
bus wires switches, say, from 1 to 0, the adjacent wires could expe- 
rience a spurious voltage drop due to cross coupling through the 
capacitances 6 , between the switching wire and the adjacent wires. 



6 This is not strictly correct, but a high capacitance between two wires can be 
viewed as a small (imaginary) resistance during high frequency switching. So the 
voltage changes from one wire will be "conducted" through the capacitance to 
the other wire, especially when the capacitance is high. 



156 Chapter 9 

The cross coupling effects were not too important in older chip 
technologies, but with the more advanced technologies the wire 
spacing was much smaller, and the capacitances between the wires 
went up to the point that cross coupling could not be ignored. 

The engineer from VLSI Technology did not believe my cross 
coupling theory at first, but after going over the capacitance values 
for the output bus wires with me, he agreed that it was plausible. 

What could be done to avoid the cross coupling problem? We 
could space the bus wires further apart to reduce the inter-wire 
coupling capacitance. Alternatively, shielding wires could be intro- 
duced between the bus wires. There were probably some other 
simple solutions. The software from VLSI Technology did not pro- 
vide any ready answer. One solution at my disposal was to give the 
output bus more time to settle before reading the output value. Of 
course, that would require a modification to the circuit design, and 
re-fabricating the chips. 

Anyway, feeling that all the chip problems were resolved, I 
sent an e-mail to the rest of the team summarizing my findings. 
The next morning, I received a message from Joe Hoane. When Joe 
inherited the Deep Thought H code, he added a massive amount of 
code for program self-testing. Using the self-testing code, Joe found 
a repeatable problem with the chip even after adding in the pre- 
read operation. After a long night, Joe meticulously pinned down 
the problem to something related to the en passant chess moves. 
The en passant rule is one of my least favorite rules in chess. Over 
half a year of my life was wasted fixing problems related to en pas- 
sant in one chess machine after another. It took me half a day to 
understand the precise nature of the problem. I had made a subtle 
logic design error. The chip misbehaved under certain conditions 
when Black had an en passant move. The easiest temporary solu- 
tion that I could think of was to use two chess chips 
simultaneously, one of the chips having the regular board, and the 
other having the board flipped by 180 degrees with White pieces 
becoming Black pieces, and vice versa. By comparing the two 
chips, the problems could be located on the fly and fixed in soft- 
ware. Adding the pre-read operation had already made the chess 
program far more complicated. The new board flipping operation, 
combined with the pre-read operation, probably increased the 
amount of code tenfold. Joe had about one week left to fix the soft- 
ware and find out whether there was any other showstopper before 
I was to hop on the plane to Beijing. Murray, who was creating the 



Bringing Up the Baby 157 

software for the new evaluation function hardware, also suddenly 
had some new tasks to perform. Using two chips simultaneously 
introduced several complications that had to be resolved in the 
evaluation software. 

Joe uncovered some additional problems but no new major 
one with the chip. I quickly made changes to the chip design and 
gave the engineer at VLSI Technology the new netlist. We had just 
enough time to get a new batch of chips for the match in February 
1996. 

Joe worked very long hours during the entire week, and the 
new program was still not quite working. We made a tape con- 
taining a snapshot of the software. On 18 September, I got on the 
plane with the tape and a printed circuit board with two chess 
chips on it. Joe had about sixty-four hours left to create a program 
that used the new chips to play chess. Murray stayed behind in 
case Joe needed any help. C J Tan took the trip too, to help smooth 
things over in Beijing. 

The plane trip was uneventful. I checked with Joe the morning 
after my arrival; the new program was still being worked on, but he 
had some things that he wanted me to look over. C J and I visited 
the IBM Beijing Research Lab during the day to do the preliminary 
setup. I tried to connect to the United States but the phone line only 
worked intermittently, and it was impossible to retrieve anything. 
The Beijing Lab was in a newly developed area that only a few years 
back was farmland. In fact there was still farmland nearby at the 
time of my visit. Apparently, the rapid economic development in 
Beijing had put serious strain on the phone line capacity, and the 
service in the outlying areas of the city suffered as a result. 

After I got back to the hotel, I checked with Joe on the status 
of the new program. The match was to take place the next after- 
noon. Joe had something that almost worked, and he believed that 
he would be able to send a playing program to me in time for the 
match, via the IBM internal network. The next morning, an 
employee from IBM China Headquarters brought a floppy disk to 
the playing site. There were four hours left to the match. The pro- 
gram on the floppy disk was loaded onto the workstation hard 
disk, and I did some simple tests. Yes, the program appeared to be 
working. Joe had put in a heroic effort. Well, there was not much 
to do but wait and pray. 

Xie Jun, a petite young lady in her early twenties, used to play 
Chinese chess before she received Chinese government sponsored 



158 Chapter 9 

training for "western" chess. She then burst onto the world chess 
scene, and surprised the chess establishment by winning the 
Women's World Chess Championship, becoming the first Asian to 
do so. As our match was a friendly one, Xie Jun got to choose the 
time control. I don't have a record of what time control was used, 
but it was probably twenty minutes for the game, with an addi- 
tional twenty seconds after each move. No game score was kept 
which gave Xie Jun an opportunity to try out new opening ideas 
without worrying about revealing her secrets. We agreed before- 
hand to play two games. The audience sat in a nearby room linked 
to the playing room via a closed circuit video hookup. Xie Jun's 
coach gave a running commentary to the audience. 

In the first game, Xie Jun apparently tried out one of her new 
openings. The 4-hour-old program got into trouble pretty quickly 
after the opening phase. On one of her moves, Xie Jun slowed 
down and studied the position. At this point the new program had 
a big score drop. There was a killer move for Xie Jun but she did 
not see it, and played a move that should have been good enough 
to win. It was, however, interesting to observe that she slowed 
down at the precise moment when there was a kill. After a few 
more moves, the new program crashed. There was an en passant 
possibility for Black that somehow crashed the program. When I 
tried to restart the program, I accidentally wiped out the internal 
score file that it kept. Xie Jun graciously agreed to play a new game. 
Precisely the same game got played. Once again, Xie Jun slowed 
down at the critical position. She must have sensed that there was 
a better move that she could play. With the short time control, she 
again missed the killer move, but her move was good enough. The 
program crashed again at the same point 7 . This time 1 somehow 
managed to get the program restarted. We played on a little while, 
then the program crashed again in a similar fashion. Since the pro- 
gram's position was getting pretty bad, we conceded the first game 
without trying to restart. 

After the first game, Xie Jun mentioned that she must have 
missed a killer move in the critical position. 1 confirmed her suspi- 
cion and told her what the program had calculated it to be. After a 
short break, we started the second game. This time the new pro- 
gram played without any crashes. Xie Jun fell into a tactical bind, 

7 We never found the bug that caused the crashes in this version of the program. 
By the time that I got back from China, Joe had the program working properly. 



Bringing Up the Baby 1 59 

and resigned as soon as the position became hopeless. I was 
relieved that she did not play on. The program could very well 
crash again if the game went on much longer. 

So the Beijing match was tied at one to one, and our match 
date with Garry Kasparov was still on. 

"No Computer Will Ever Beat Me" 

When I was at Carnegie Mellon, the graduate student organization 
in the Computer Science Department operated a Coke machine 
that dispensed cheap bottled Coke for thirty-five cents a bottle. The 
proceeds were the main funding source for many of the TG parties 
in the Department. About a year or two into my stay at Carnegie 
Mellon, some graduate students and engineering staff got together 
and decided to make an electronic interface for the Coke machine. 
It then became possible to "finger" it electronically from the con- 
fines of one's office, and find out whether any cold bottles were left 
in the machine. In fact, you could finger the Coke machine in 
Pittsburgh, Pennsylvania from, say, Palo Alto, California, although 
this capability probably offered very little utility. 

The Coke machine was located on the third floor of my build- 
ing. My office was on the eighth floor so I fingered it a lot, until I 
found an alternative supply — one of the bigger offices on my floor 
had a large refrigerator that was used as an auxiliary Coke machine 
by a few students, who, in exchange for the privilege, would help 
load up the refrigerator from the storage room on the 3rd floor. 

Jay Sipelstein was an occupant of this big office. Jay had 
received his Bachelor degree from Yale University and while he was 
there, he used to hang around the Yale Chess Club. One day, Jay 
mentioned to me that there was a Yale student — an International 
Master — who would give simultaneous exhibitions in the chess 
club and beat everybody. He said that someone had asked the 
International Master whether he thought any computer would 
ever beat him. The International Master answered, "No computer 
will ever beat me." At the time he made the boast, he probably had 
good reasons to be smug. A few years before, he had been on a tel- 
evision show slaughtering Cray Blitz in blitz games. He was also an 
up-and-comer. The year after the International Master graduated 
from Yale, he became a Grandmaster. The name of this new 
Grandmaster was Joel Benjamin. 

The tied match against Xie Jun in Beijing kept the match against 



160 Chapter 9 

Garry Kasparov alive. Xie Jun was a Grandmaster, but our friendly 
exhibition was played at a shortened time control, and we wanted 
to play some test matches under more serious conditions. 

We did not have the full power of Deep Blue, which would be 
about thirty times faster than the machine we had, so playing at 
an even level with Grandmasters at a regulation tournament time 
control would be a satisfactory result. We scheduled to play in our 
lab, two games each, against three leading US Grandmasters. The 
Grandmasters were offered incentives to try to win the match or, 
failing that, to try to draw the match. We were also using the test 
matches to screen for a Grandmaster to help with the opening 
preparation for the match against Garry Kasparov, but none of the 
Grandmasters knew about the job opening. We always knew that 
we would need a Grandmaster on the team in the final match. Of 
course, the Hong Kong incident had also made that crystal clear. 

Deep Blue Jr, the scaled down machine, won one of the 
matches, drew another, and lost one. 

The winning Grandmaster was none other than Joel 
Benjamin. Deep Blue Jr was in the very early stages of its develop- 
ment, so losing one of the matches was not a big surprise. The way 
Joel won the match 2-0, however, was a big surprise. 

The first game against Joel was a typical computer affair. Deep 
Blue Jr misjudged the endgame position, and Joel's positional 
judgment proved superior. We knew that there was lots of room for 
improvement in Deep Blue Jr's evaluation function, so there was 
nothing to be alarmed about. 

The second game was a long one. It started in the afternoon 
and finished late at night. When the endgame phase was reached, 
it was clear that Joel was better but the position was probably 
drawn. The incentives offered were for winning or drawing the 
match, so the practical thing for Joel to do was to offer a draw, win 
the match, take the cheque, and say thank you. Joel pressed on. 
Murray and I were getting hungry by now. For a short while, it 
appeared that Joel was not making progress. But somehow he slow- 
ly and patiently tricked Deep Blue Jr into making one slight 
mistake after another, until he got complete control of the situa- 
tion. He then pressured Deep Blue Jr to finally give up material, 
and shortly afterwards we resigned for the program. 

After walking Joel out of the building, Murray and I were 
silent for a while. There was a lot of work ahead, but we also both 
knew that we had found the person to do the opening preparation 



Bringing Up the Baby 161 

for Deep Blue. Joel did not know it yet, but we had found the last 
piece of the puzzle. The question was whether we had enough time 
to create something that could beat Garry Kasparov in February 
1996. 



ft) 



CHAPTER 10 



A Living Mount Everest 



At a Seaside Resort 

I became aware of the ACM's interest in hosting a Kasparov vs. 
Deep Blue match under unusual circumstances, well before the 
Hong Kong event. 

While I was working on the Deep Blue chess chip, I consid- 
ered attending computer chess events a nuisance. Even if I had not 
been involved with the preparation for the event, I knew perfectly 
well that I would not be able to do any serious work on the chess 
chip for a few weeks — I could not stop myself from following our 
progress in the event. 

When the 1994 ACM Computer Chess Championship came 
along, I was against going. The chess chip design work was enter- 
ing the critical final phase, and given that the entire project hinged 
on the chip working, it seemed foolhardy to waste time on a tour- 
nament, albeit an important one. Monty Newborn, the organizer 
of the ACM event, and a good friend of C J Tan, persuaded him to 
enter Deep Thought 11 into the Championship, and I was over- 
ruled. 

It turned out that both Murray and Joe had already scheduled 
their vacations around the time of the championship. Joe might be 
able to show up for the closing part of the championship, but 
would definitely miss the early part. I was the only person who 
could go to the ACM event and operate the program. 

The championship was held in Cape May, a seaside resort 
town at the southern tip of New Jersey, about a four-hour drive 
from our lab in Hawthorne. Since there was not much I could do 
to avoid the "obligation" to go, I treated the trip as a vacation from 
the chip design work 1 . In that sense, the drive was quite pleas- 

162 



A Living Mount Everest 163 

ant — a scenic route, sun, fresh air, and lots of sea breeze. At the 
hotel, I bumped into several members of the other teams. 
Everyone was trying to feel each other out. Who would be danger- 
ous this year? Star Socrates from MIT was attracting a lot of 
attention, as it would be the first program since Deep Thought 11 
to cross over the one million positions per second mark. 

On the first day of the tournament, the weather turned 
stormy. What I did not know was that the storm was much worse 
in Hawthorne. Our first round meeting was with Zarkov, a 
respectable opponent. Prior to the tournament, Murray had been 
experimenting with the King's Gambit opening. The King's 
Gambit is a violent opening dating back to the romantic age of 
chess in the 19th century. It is rarely seen these days at top-level 
chess events. It is also probably not a sound opening for White. 
Murray had found some opening innovations for White in the 
King's Gambit with the help of Deep Thought II, and wanted to try 
them out. Zarkov, however, was not cooperating. It played a line 
that led to a position which Murray had not examined and it was 
a good, possibly even a winning, position for Black. Deep Thought 
II was not too happy about the situation it found itself in, but there 
was no obvious kill for Black. All the other teams who came by our 
table thought that we were lost. To tell the truth, I was not too con- 
cerned, but I should have been. Ignorance is bliss. Had he been 
present, Murray would have been blaming himself. I had seen 
Deep Thought II win brilliantly using the King's Gambit, even 
though it usually did not like White's position out of the opening. 
I just did not know that the position on the chessboard gave us no 
counter chance, if Zarkov had known how to play it correctly. 
With Zarkov's unwitting help, Deep Thought II seized enough 
opportunities, and eventually John Stanback, the author of Zarkov, 
resigned for Zarkov as soon as the position became hopeless. 
Neither he nor I knew at the time that if the game had lasted for 
another half-hour, Deep Thought II would have had to forfeit. 

While our game against Zarkov was going on, a sudden down- 
pour took place in Hawthorne. Over two inches of rain fell within 
an hour, and the buildup of water shorted the transformer provid- 



1 With the heavy pressure to complete the chip, I had not taken any real vacation 
days for years by then. IBM did not allow their employees to accumulate vaca- 
tion days from one year to another. I ended up going to work on "vacation" 
days. It was a sacrifice that I will not make again. 



164 Chapter 10 

ing power to the lab. For the last half-hour of the game against 
Zarkov, Deep Thought II was using the battery power of our UPS 
(Uninterruptible Power Supply) backup. The battery could last 
only an hour or so, and it ran out of power by the second round. I 
found that I could not connect to Deep Thought II at all, so I asked 
for a time-out from the tournament director to investigate what 
was going on. I phoned Jerry Brody, our support engineer 2 , and 
the only person in the group not out of town, and asked him to get 
into the lab to find out what went wrong. Twenty minutes later, 
Jerry gave me the bad news. There was no power in the lab, and 
there was no estimate about when it would be back. With the con- 
sent of our opponent, the tournament director granted us more 
time. In computer chess events, there is usually a provision for 
unlimited communication time-out. Power failure was never 
explicitly covered in the rules, although an argument could be 
made that it was equivalent to a communication failure. By eleven 
pm, we still could not get the machine back. Our opponent was 
asked whether the game could be rescheduled to the next day; he 
declined and we forfeited the game without playing a single move. 
The machine was up and running at about one am, but it was too 
late. (Eighteen months later, during the first Kasparov vs. Deep 
Blue match, we informed the electric company about our power 
requirements. After seeing the front-page news that the match was 
making, the company made special arrangements to ensure a con- 
tinuous supply of electric power to IBM Research Lab. They 
certainly did not like the prospect of getting worldwide notoriety 
for causing Deep Blue to lose a game due to a power failure on their 
part!) 

Back to the tournament. The event was a five-round Swiss. For 
all practical purposes, losing one round by forfeit pretty much 
eliminated our chance of winning the tournament. The best we 
could do would be 4 points out of 5. Based on past experience, this 
was usually insufficient to win a tournament like this one. I was 
feeling miserable. 

Our third round opponent was WChess which put up a stiff 
fight, but Deep Thought II won without too much trouble. The 
fourth round opponent was Star Socrates, whom we would face 

2 Jerry had been with the team ever since we joined IBM. He did not really get 
involved with the technical work of Deep Blue, but was more of a provider of 
general support, including parts acquisition and order/contract processing. 



A Living Mount Everest 165 

again a year later in Hong Kong. During the game, the MIT team 
was very pleased to observe that Star Socrates was reporting 
"greater" search depth than Deep Thought II. At this stage of the 
tournament, Star Socrates was the only team with a perfect score. 
It also had White, so the MIT team must have felt that they had a 
great chance to win the event. Just as in Hong Kong a year later, 
Deep Thought II won the game relatively smoothly, even though 
there were anxious moments in both games. After all, anything 
that searches over a million positions per second is not something 
to be taken lightly, and Star Socrates was definitely a contender. 

The fifth round pairing became very interesting. The only 
program that remained undefeated was Mchess, the second round 
opponent against whom we had forfeited. Mchess had conceded 
only one draw. Normally, in a Swiss tournament, you would not 
play any of your opponents twice in the same tournament. But 
Mchess and Deep Thought II never did really play each other, and 
Deep Thought II had already played all the other top programs, 
with the sole exception of Mchess. So Deep Thought II played 
Mchess again, this time for real, and won the game fairly quickly. 
The tournament now became a question of whether Star Socrates 
could win its last game against Zarkov, in which case it would tie 
for first. Star Socrates lost, and Deep Thought II won the title clean- 
ly, despite the second round forfeit. 

The next morning, I had a good breakfast in the hotel. The 
sun was shining. The beach looked enticing, and Deep Thought II 
had pulled off an impossible feat. I was in a very good mood. 
Monty Newborn, the tournament organizer, asked whether he and 
his friend could join me. Monty introduced his friend as Frank 
Friedman, a professor from Temple University near Philadelphia. 
Frank was also Monty's boss in the ACM organization. Monty and 
Frank asked me when the new chess chip would be ready, and I 
told them that it should be available some time the following year. 
Monty then broached the real subject. He asked, "Would IBM be 
interested in playing a match with Garry Kasparov during the ACM 
Annual Conference in 1996?" This was getting interesting. I 
answered, "I don't speak for IBM, but I would imagine so." Then I 
added, "The match would be quite expensive." Monty said that he 
expected the prize fund for the match to be something like 
$300,000 to $400,000. I told him 1 believed that the prize fund 
would have to be much higher, because the match would have 
greater historical importance than a regular World Chess 



166 Chapter 10 

Championship match. The last World Chess Championship match 
had a prize fund of well over one million dollars. I told Monty and 
Frank, "I would be very surprised if Garry settles for anything less 
than a million dollars." Monty had some ideas of his own, "Well, 
if we only play, say, six games, then we could get the prize fund 
down." "Perhaps", I answered without much conviction. We con- 
cluded the conversation without making any real commitment. 
ACM would start talking with Garry and get his feedback. I placed 
their odds of getting Kasparov on board at 1 in 5. 

The ACM people proved me wrong. Garry eventually agreed 
to play a six-game match against Deep Blue in Philadelphia for a 
total prize fund of $500,000, with $400,000 to the winner, and 
$100,000 to the loser. Garry was so confident of winning the 
match that he agreed to it without demanding more money. In 
fact, he wanted the match to be all or nothing with the loser get- 
ting nothing but was eventually talked out of the idea. IBM 
Research put up the prize fund. Deep Blue's share of the prize 
would go back to IBM to fund more research, not necessarily in 
computer chess. ACM was scheduling the match as part of its 50th 
anniversary celebration. 1996 was also the 50th anniversary for 
ENIAC, the first electronic computer, and ENIAC was built near 
Philadelphia. Therefore, ACM decided to kick off its year-long cel- 
ebration in Philadelphia, and the Kasparov vs. Deep Blue match 
(10-17 February 1996) was the first event on the agenda. The 
match was to be played at the rate of one game per day, with one 
rest day following every two game days. The game days were 10, 
11, 13, 14, 16 and 17 February. 

Winter in Philadelphia 

In December 1995, two months before the Philadelphia match, we 
received a small batch of the revised chess chips from VLSI 
Technology. The chips worked properly this time. There was no 
more need to do the pre-reading operation. There were no more 
problems with en passant, and no more chip mirroring. 
Grandmaster Joel Benjamin agreed to be Deep Blue's match sec- 
ond. The time remaining until the match was so short that his 
main function was limited to preparing the opening moves. Since 
most of the chips were still in VLSI's testing and packaging lines, 
Deep Blue did not really exist yet, and Joel did pretty much all his 
opening preparation on Deep Blue Jr, the scaled down workstation 
version of Deep Blue. 



A Living Mount Everest 167 

About two weeks before the match, the remainder of the chips 
arrived. It took the better half of a week to test them all out, and 
then two days or so to load them up into the IBM RS/6000 SP 
supercomputer, that was to be Deep Blue's host. We had barely 
enough time before the match to test out all the hardware, and do 
the basic software testing on the big machine. We also did not 
have as many chess chips as we would have liked. 

Given our experience in Hong Kong, we decided to set up two 
backup machines at the playing site in Philadelphia, with the main 
RS/6000 SP supercomputer staying behind at the lab in Yorktown 
Heights, New York 3 . One backup machine was a regular IBM 
RS/6000 workstation running the Deep Bluejr program. The other 
machine was a small four-way RS/6000 SP supercomputer, with 
four workstation nodes. The main machine was a thirty-six-way 4 
RS/6000 SP supercomputer. Theoretically, with a thirty-six-way 
machine, the system could have up to 576 chess chips working 
together, with sixteen chess chips per workstation node. Due to the 
limited number of chess chips available, each node on the main 
machine only had six chips, for an aggregate of 216 chess chips 
total. The chess chips in this batch searched about 1.6 million posi- 
tions per second, and the theoretical maximum search speed of 
this first Deep Blue was about 300 million positions per second. 
The observed search speed was about 100 million positions per 
second. The software overhead associated with running the search 
in parallel was quite high. We probably could have speeded up the 
machine by up to a factor of two, if we had had another six 
months to work on the software. 

The chess chips evaluated the chess positions in far more 
detail than possibly any other chess program that had ever existed. 
Almost all the evaluation function terms were computed directly 
on chip s . As it was, to perform the same computation on a gener- 
al purpose computer, as that done by the 1996 Deep Blue, would 

3 We had moved from Hawthorne, New York to the main IBM Research lab in 
Yorktown Heights, New York. 

4 This is the real number. At the time of the match, C J Tan gave out the number 
of host CPUs as 32. We were instructed to stick to the number 32 to avoid con- 
fusion. The 36-way machine consisted of two 16-CPU frames plus four 
workstations. 



5 



The software evaluation function assigned weights for the evaluation terms, 
instead of computing the evaluation function itself as in typical chess programs. 



168 Chapter 10 

need at least a one trillion instructions per second general-purpose 
system 6 . It was the most powerful chess machine ever built. It had 
also never played a single chess game going into the Philadelphia 
match. 

Philadelphia was sufficiently close that driving there made 
more sense than taking a commercial flight. On 8 February 1996, 
Murray, Joe, Jerry Brody, and I finished the last pre-match testing 
and drove to Philadelphia while Grandmaster Joel Benjamin took 
the train. I don't know how Joel felt during his trip but in our car, 
the atmosphere was a little bit somber. We had come so far, and yet 
we were still a long way from where we wanted to be. Deep Blue 
was still a two-week old baby. It was a very powerful baby, but a 
baby nonetheless. Would it be the baby Hercules that strangled the 
two serpents sent by the Goddess Hera? Or were we sending a help- 
less baby up as a tribute to placate the sea monster Cetus, but 
without the aid of Perseus? We were afraid that it would be the lat- 
ter. 

We arrived at the Philadelphia Marriott Hotel in the late after- 
noon. The conference and the match were to be held in the 
Philadelphia Convention Center right across the street. An 
enclosed walking bridge connected the hotel directly to the center. 
The entrance hall at the convention center end of the bridge was 
decorated with several high towers that reminded me of chess 
rooks. While checking into the hotel, we were greeted by Terrie 
Phoenix, the ACM director of public relations, and Marcy Holle, 
the communications person from IBM Research in charge of press 
relations for the chess project. Marcy had replaced Jerry Present a 
few months back when he retired. Terrie and Marcy were on their 
way somewhere else, but they informed us about the press confer- 
ence to be held the next day. 

A small contingent of IBMers had arrived several days earlier 
to deal with all the logistics, video setup, web content distribution, 
and so on. We met with some of them for dinner. They told us 
some of the surprises that they encountered. 

When the match contract was signed, it was decided that the 
match would be played in a large theater in the convention center, 



6 Put it in another way. Each chess chip was roughly equivalent to a multi-million- 
dollar general purpose supercomputer. In particular, a single chess chip was 
actually more powerful for chess processing than the host RS/6»00 SP super- 
computer we used for Deep Blue. 



A Living Mount Everest 169 

to accommodate spectators who wanted to see the match directly. 
However, the person in charge of making the arrangements left 
ACM before the deposit had been put down. By the time ACM got 
back to it, the theater had already been rented out. A lecture room 
was found to serve as the match room, with an adjoining room 
reserved as Garry's dressing room. Matt Thoennes was the IBMer in 
charge of making the whole arrangement work. A raised stage was 
set up in the lecture room, and several video cameras were brought 
in to provide video feeds to the audience. There was enough space 
for about seven or eight rows of spectator seats. When Garry's 
mother, Clara Kasparova, inspected the match site, she objected to 
the idea of having rows of spectators. A partitioning wall was 
added leaving just one row of seats for Garry's entourage, ACM 
officials, and IBM personnel. The paying audience would stay in 
the commentary room. Another interesting request by Clara, at 
least from Matt's point of view, was for a port-a-john (portable toi- 
let) in Garry's dressing room. Matt asked someone at the ACM how 
to get a port-a-john, and was surprised to receive a catalog of all 
sorts of them, from the most basic to fancy executive models with 
luxurious amenities. Besides the port-a-john, food and drink 
(including Toblerone chocolate, apparently one of Garry's 
favorites) were provided. 

The morning after our arrival, while we were having breakfast 
in the hotel restaurant, Clara and Garry's coach Yuri Dohokian 
were at another table. It was the first time that I had seen either of 
them. Yuri was stout looking and young. Clara was elegant in her 
demeanor, but you could sense that here was a woman who had 
seen both good times and bad. It took great inner strength for a 
young widow to raise Garry to the great success that he became. 
Garry is a strong-willed person, and Clara probably had a strong 
hand in shaping his personality. Clara was friendly, but she also 
commanded respect. Throughout Garry's career, whenever there 
was a big match situation, she was always there to lend her sup- 
port. For the Philadelphia match, Clara proved to have a calming 
influence on Garry. 

The press turnout at the pre-match press conference was 
decent, with around twenty reporters showing up. The playing 
color was decided with Deep Blue having White in the first game. 
At the press conference Garry predicted a win for himself but did 
not guess at the final score. Most were predicting an easy win for 
Garry; David Levy went so far as to predict a 6-0 wipeout win. 



170 Chapter 10 

Privately, I believed that David's result was highly improbable, but 
it was not out of the question that we could lose the match very 
badly. No one outside IBM knew at the time that our machine was 
only two weeks old. We were nowhere near ready for the match. 

The first match game began with a lot of anxiety for the Deep 
Blue team. We had never seen Deep Blue play a single game our- 
selves, and had no idea what to expect. Compounding this, we had 
some early glitches in starting the game owing to computer net- 
work problems in the Yorktown Lab. 

At Carnegie Mellon, during public games we isolated the Deep 
Thought workstation from the departmental local network to 
avoid the occasional network outage, which could render the 
machine unusable. Originally, we also planned to isolate the 
RS/6000 SI J running Deep Blue from the rest of the Research lab 
local network. But the plan was not workable — the machine was 
semi-public for the entire lab, and it could not be isolated because 
of the way that it was set up. As a result of the occasional network 
problems, Deep Blue took a little while to start up for the first 
game. Mike Valvo, the match arbiter, started Deep Blue's clock 
while we were still waiting for Deep Blue to start up. 

Deep Blue opened the first game with e4, moving the pawn 
in front of the king by two squares. Garry responded with c5, 
moving the queen bishop pawn forward two squares. This was 
the Sicilian defense, Garry's favorite defense against e4. Before 
the match, we had some debate about whether he would play it. 
Joel's opinion was that Garry was very disciplined in opening 
selection and he fully expected him to play the Sicilian. There 
were many good move choices for White after Garry's c5 but, 
unfortunately, most of them had reams of associated analysis. 
Joel only had months to prepare Deep Blue's opening book, so 
we chose to play less traveled opening lines. Deep Blue answered 
c3, moving the queen bishop pawn forward one square. This is 
a sound, but somewhat irregular, opening for White. The advan- 
tage was that the c3 Sicilian was rarely played at the 
Grandmaster level and above 7 , and as a result there was no need 
to do extensive opening preparation. The disadvantage was that 
the opening was a little bit dry, and did not seem to offer much 
in terms of winning chances for White. At least, that was my 

7 The c3 Sicilian was actually quite common in computer chess play. Most pro- 
grammers did not have a lot of time to prepare their opening books. 



A Living Mount Everest 1 71 

opinion based on my past experience with Deep Thought and 
Deep Thought II, playing both sides of the c3 Sicilian. Joel had 
a different opinion. 

The c3 Sicilian frequently leads to the so-called "isolated queen 
pawn" position, where White's queen pawn ends up with no friend- 
ly pawn on the adjacent files. The isolated queen pawn is usually a 
handicap in endgames, but in middle games it offers dynamic 
attacking chances for White. The problem was that most chess pro- 
grams were usually good reactionary defenders, but not good 
proactive attackers. Joel spent a significant amount of time coach- 
ing Deep Blue Jr how to attack from the positions arising out of the 
c3 Sicilian. There was one nagging problem. Deep Blue Jr did not 
like White's positions, and assessed the position to be better for 
Black. Usually, Deep Blue would still play the right attacking moves, 
but occasionally it would allow Black to repeat the position since it 
expected Black to avoid the repetition draw. What happened was 
that the 1996 version of Deep Blue did not have full grasp of the 
idea of attacking potential, and hence made the wrong assessment. 

The fact that Deep Blue did not expect Black to repeat the 
position created an intriguing situation in the first game. On move 
ten, Garry produced a new move that took Deep Blue out of its 
opening book. Three moves later, Deep Blue presented Garry with 
an unexpected choice. When I played move 13, NbS for Deep Blue 
attacking his queen, he was taken aback. 

Garry looked at the board intently for a moment, then raised 
his eyebrows and stole a glance at me. I could not help but smile 
momentarily on seeing his expression. Deep Blue did not really 
"enjoy" its position, and yet Garry was taking the position very 
seriously. My smile faded the moment I realized that I was smiling. 
I don't know how he interpreted my fleeting smile but he spent a 
long time on his reply. One of his choices was to undo his last 
move, Qd6, by moving his queen back to d8. In this case, Deep 
Blue would have undone its own last move by playing Nc3 and 
repeating the position, effectively offering a draw. Deep Blue 
played 13. Nb5, in part because it was willing to settle for a draw. 
Garry did not move his queen back to d8, either he did not want 
an early draw, or he did not realize that Deep Blue would have 
taken an early draw. Garry finally placed his queen on e7, a square 
that Deep Blue considered somewhat awkward, as one of Garry's 
knights was soon pinned against his queen. From then on, his 
position steadily went downhill and became desperate. 



172 Chapter 10 



8 

7 
6 

5 
4 

3 
2 

1 




V'>2>'J 



Hi 



■Mill 



. t jt^ l 



Y M 






A.Jm I 



p 

B, IH 



M9&S 



> ^W^ 



_rf 



abcdefgh 

Position after 13. NbS. 

From there, Garry proceeded to whip up a nasty looking 
attack that surely would have unnerved any carbon-based chess 
player. Deep Blue totally ignored his attack, and merrily snatched 
up queenside pawns. Garry's attack, meanwhile, was looking more 
dangerous as time went on. All that he needed was one free move, 
and the game would be all over in his favor. At this point, Deep 
Blue unleashed its own attack. Garry could beat back the attack but 
would lose most of his material and see his attack dissipated. For 
the last few moves of the game, he knew that his position was 
hopeless, and yet Deep Blue, with over an hour of time left on its 
clock, was still taking time to calculate its best move. After Deep 
Blue's move 37, Garry had had enough, offered his hand, and 
resigned. Garry had about five minutes left on his clock for his 
next four moves, while Deep Blue had about an hour left for its 
next three moves. It was the first win ever by a computer over a 
World Chess Champion under regulation time control. History 
was made! 

After shaking hands with me, the first question that Garry 
asked was, "Why did it take so long to play its moves? A PC pro- 
gram would have seen that it was winning instantaneously." I 
shrugged my shoulders and told him that Deep Blue simply had 
too much time left on its clock. Next, he asked with a pained look, 
"Where could I have played better?" This question took me by sur- 



A Living Mount Everest 1 73 

prise. I could barely play chess and the greatest chess mind ever was 
asking me where he could have improved his play. Of course, 
Garry's question really was "Where did Deep Blue think that I 
could have played better?" I tried to answer his question as best as 
I could from what I remembered of Deep Blue's analysis. Sensing 
that I was not seeing the game on the same level as he and Deep 
Blue, Garry gave up after a few minutes, and quickly left the build- 
ing without talking to anyone outside of his entourage. 

It did not really hit me how momentous the occasion was 
until Joel Benjamin came up to the stage to shake my hand. Joel 
said, with a broad smile, "I dreamt about beating Kasparov myself 
and shaking his hand," adding, "This is almost as good." 
Throughout the entire game I had been reining in my emotion. I 
now felt both elated and relieved. Murray and Joe were ecstatic as 
well. A reporter came by to get a quick quote and then rushed out 
to file the copy. There weren't that many reporters around then, 
but that changed very quickly in the next few days. 

The IBM match web site was swamped during the first game. 
The newspaper USA Today, carried a front page article stating that 
IBM's Deep Blue was doing well but the web site was not. 
Apparently the web master had thought that only a few thousand 
surfers would be visiting the site during the game. Had we been 
asked, we could have told the web master to expect at least tens of 
thousands visitors and possibly hundreds of thousands 8 . One of 
the web visitors was Jonathan Schaeffer, trying to connect from the 
University of Alberta in Edmonton, Canada. Unable to get 
through, Jonathan scanned the news service and found a Reuters 
story erroneously saying that Garry was White. Our friend Dap 
Hartmann, was in Boston and had problems connecting as well. 
He and Jonathan agreed to both try and access the IBM site peri- 
odically and keep each other posted. They were able to follow the 
game by getting the board position from time to time. The web dis- 
play for the board position did not state which player had what 
color so Jonathan and Dap "watched" Garry score a nice win as 
White. As predicted. End of story. It was twenty-four hours later 
that Jonathan found out that Garry had lost. Jonathan's situation 
was probably quite typical for a lot of web visitors on the first day. 

8 Contrary to the web master's initial expectation, the IBM match web site turned 
into the most popular web site up to that point in time, exceeding even the 
Superbowl web site and the swim suite edition web site of Sports Illustrated. 



174 Chapter 10 

IBM started mobilizing the web team during the first game. An 
IBM RS/6000 SP supercomputer, the same type of machine as Deep 
Blue's host computer, was enlisted as the new web server to handle 
the workload. After game 2, USA Today carried another front page 
article, this time stating that IBM had fixed the web site. 

That night after game one, the IBM team had a celebration 
dinner. No matter what happened in the remainder of the match, 
we had made our point. The Kasparov camp apparently was in a 
crisis. According to Frederic Friedel, who was serving as Garry's 
computer consultant, Garry went for a late night walk on the 
Philadelphia streets in below freezing temperatures. During the 
walk, he asked, "Frederic, what if this thing is invincible?" 

Quality from Quantity 

Joel and Murray worked together on Deep Blue's opening moves 
for game two during the night and the next morning. 1 was pres- 
ent during part of the opening preparation session. Test positions 
related to the opening were entered into a test file by hand, and 
Deep Blue analyzed the positions in the test file overnight for 
human analysis the next morning. We all agreed that if there was 
ever going to be another match, we would need to automate the 
opening preparation process, both to speed it up, and to reduce the 
chance for human error. We had no idea how prophetic our obser- 
vation was. After the opening preparation session, Murray 
uploaded the opening file, which was prepared on his IBM 
Thinkpad laptop computer, to Deep Blue in our Yorktown lab. Or 
so he thought. 

Game two started innocently enough. On the second move, 
Deep Blue was out of its opening book. I did not know it at the 
time, but something had gone wrong when the opening file was 
uploaded, and Deep Blue did not have the new opening book. 
Since we never did use the new opening book, the opening pre- 
pared for game two ended up being played in game four. Deep Blue 
was only out of the opening book superficially, however. Ever since 
Deep Thought II, Murray had created what we called the "extend- 
ed book", which tried to capture the human concept of "opening 
theory". When Deep Blue was out of the opening book in the tra- 
ditional sense, it started to use statistics from games played by 
Grandmasters to help it decide which move to play. If a move in a 
particular position had been played by strong Grandmasters with 
great results, then the move would be given a bonus, proportional 



A Living Mount Everest 1 75 

to how good the results were. This gave Deep Blue a bias to play 
the "known good lines", but still allowed it a chance to play unex- 
pected moves that, in its assessment, appeared to give better 
chances. 

Murray's human error in transferring the file was not the first 
in this game. On move six, two pawn captures were possible for 
Deep Blue. I mistakenly played the wrong pawn capture on the 
official chessboard. Murray rushed in to inform Valvo, the arbiter, 
and me about the mistake. It was soon corrected, and time was 
added back to Garry's clock. The extended book worked quite well 
in this game. By the time that Deep Blue left known opening the- 
ory, Garry had only a slight advantage. Murray's human error, 
ironically, in combination with his earlier work on the extended 
book, produced a very playable position for Deep Blue. 

On move eighteen, Garry offered a sham pawn sacrifice and 
Deep Blue declined. Taking the pawn would have given him a 
strong attack. On the next move, he offered a real pawn sacrifice. 
Post-game analysis indicated that Garry's pawn sacrifice should 
have led to a draw if Deep Blue had played the best moves. After 
the game, while trying to analyze what went wrong, we found an 
evaluation software "bug" in Deep Blue. The Deep Blue chess 
chip contained a piece of hardware that recognized minor pieces 
(bishops and knights) on so-called "outpost" squares, squares on 
which they were protected by friendly pawn(s) and could no 
longer be attacked by the opponent's pawns. The chess chip fur- 
ther identified whether the minor pieces could no longer be 
challenged by (or exchanged with) the opponent's minor piece. 
It is well known in chess lore that if an outposted knight can no 
longer be challenged, then it can be as valuable as a rook. 
However, unchallengeable outposted bishops are perhaps no bet- 
ter than regular outposted bishops. Unfortunately, the software 
code that assigned the values, wrongly assigned unchallengeable 
bishops the same values as unchallengeable outposted knights. 
This gross error in the software evaluation plagued Deep Blue's 
play for the remainder of game two. If we had had more time for 
preparation, as we did in the Deep Blue rematch, this fairly obvi- 
ous "bug" would have been discovered during test games long 
beforehand. 

A sequence of piece exchanges ensued after Garry's pawn sac- 
rifice. Deep Blue had at least two chances for relatively easy draws 
during the sequence. The first one it missed because of its desire to 



176 Chapter 10 

park its "unchallengeable" outposted bishop. The second one it 
missed because of a bias placed into the program to avoid trading 
queens. Going into the 1996 match, we believed that Garry was 
definitely the stronger player, and to increase our tactical chances, 
we gave Deep Blue a large penalty for trading the queens. The 
queen trade avoidance bias backfired in this game. Garry, after 
observing the bias first hand, exploited it masterfully in later 
games. 

After the dust settled, Garry had a clear edge, but Deep Blue 
still had drawing chances. He then allowed his most dangerous 
pawn to be traded for one of Deep Blue's weak pawns, unnecessar- 
ily, according to Deep Blue. An ending with each side having a 
queen and a bishop, with the bishops on opposite colored squares, 
appeared on the board. The ending was probably drawn, but Deep 
Blue's position was very difficult to hold 9 . Garry initiated a long 
and subtle maneuver that netted him a pawn. With the pawn 
deficit, Deep Blue's position began to crumble. On move 73, 
Murray, who had replaced me as the operator after the first time 
control, resigned for Deep Blue. 

Garry was exultant, went immediately to the commentary 
room, and spoke to the audience for over thirty minutes. Clara sig- 
naled him to cut it short, but in his excited state he went into a 
detailed post mortem analysis of the game, and gave a summary of 
the match so far. The audience thoroughly enjoyed it. During the 
summary, he said, "I congratulate the researchers at IBM for a fan- 
tastic achievement. They have succeeded in converting quantity 
into quality." That helped reduce the sting from the loss that we 
just suffered. The next praise took us by surprise. "In certain kinds 
of position, it sees so deeply that it plays like God." 

That night, we went over the game carefully trying to find out 
what went wrong. When Murray discovered the problem with the 
outposted bishop, we were all shocked. The surprise was not from 
the fact that we had a serious bug — we expected many more bugs, 
given the little time that we had to prepare for the match. The big 
surprise was how close we had been to possibly winning the 
match. Without the bug, Deep Blue might very well have drawn 
game two. Garry already had lots of self-doubt after game one. A 

9 After the game, Garry showed a possible drawing line that entailed Deep Blue 
sacrificing a pawn immediately. Deep Blue did not understand his drawing line 
at all. 



A Living Mount Everest 1 77 

drawn game two could have tilted the match odds heavily in our 
favor, not just because we would remain up by one game, but also 
because of the psychological pressure that would be placed on 
Garry. 

The next day, February 12, was in theory a "rest" day 
although we, and probably they, were hard at work. Two members 
of Garry's camp did get some rest. Terrie Phoenix took Clara 
Kasparova and Frederic Friedel shopping. During the 1996 match, 
the two teams maintained a cordial relationship, and Terrie proba- 
bly played the major role in smoothing things along. Terrie left 
ACM before the 1997 rematch. The relationship between the teams 
was more competitive at the 1997 rematch, in part because of the 
very real possibility of Deep Blue winning the match, but the 
absence of someone like Terrie most likely played a part as well. 

The Deep Blue team quickly settled into a routine for this rest 
day. Murray and Joel had the most stressful work, preparing the 
opening for the next two games. I spent my time running tests, 
which sometimes revealed problems in the chess program. The 
program itself was sufficiently stable, and most of the problems 
were endgame related. Joe acted beyond the call of duty as the fire 
fighter who fixed up the problems as they came along. He had 
another task during rest days — speeding up the program. The par- 
allel chess program on the IBM RS/6000 SI' was quite a bit more 
complicated than Deep Blue Jr, which ran on a single RS/6000 
workstation. It was also a fairly new program, and there was lots of 
room for improvement. Some of the improvement was done 
throughout the match, but most of it came during the rest days. 

Once in a while, I would remove myself from the testing and 
join Murray and Joel in their discussions about our opening choic- 
es. Deep Blue won the first game, but neither Joel nor Murray was 
happy with how the opening went; Garry could have had an easy 
repetition draw. In a short match, you have to try to win every 
time that you have White. We needed something better. Murray 
and Joel worked late into the night on the opening problems. Joe 
also stayed late on his own task. I decided to take a break and 
joined the rest of the people from IBM for a nice dinner. 1 like to 
eat, and no work was going to stop me from enjoying a good meal. 

Murray and Joel did not come up with a good solution to our 
opening problems during the rest day. But the morning before the 
third game of the match, Joel declared that he had had an inspira- 
tion during the night. Deep Blue would deviate early from the 



178 Chapter 10 

moves it played in game one. There were several possible continu- 
ations that could be played by Garry. We checked out a few, and 
they looked good for Deep Blue. In some of them, Deep Blue found 
clearly winning lines for itself. This was looking great! There was 
not enough time to check everything out but we did as much as 
we could and put in the new opening lines. 

Everything went as expected; Garry played into our prepared 
opening lines. After move twelve, Deep Blue was out of our 
checked book lines, but was still in one of our expected sequences. 
On move eighteen, Joel expected Deep Blue to play 18. Be5, and 
set up a potential attack. Deep Blue refused to play 18. Be5, and 
moved one of its rooks instead. It turned out that Garry had a 
strong refutation against 18. Be5, and he had seen the refutation 
several moves back at the board. Without 18. Be5, Deep Blue had 
no attack but several weak pawns that had to be defended imme- 
diately, or at least that was the way it looked. While Deep Blue's 
position looked optically grim, it was an illusion. After several 
inhuman looking moves, it became clear that Deep Blue was not 
losing, and might even be slightly ahead. On move 39, both sides 
agreed that there was no point in playing on, and a draw was 
agreed. It was an impressive game by both Garry and Deep Blue, 
but we had wasted one of our Whites. 

Since the opening prepared for game two was never played, 
there was no opening preparation session in the morning before 
game four. This time, Deep Blue played a Slav defense. Deep Blue's 
position after leaving its opening book turned out to be less favor- 
able than the one it reached on its own in game two. On move 
fifteen, Garry had a chance to make a speculative sacrifice, but 
decided against it. After the match, Garry visited IBM Research and 
analyzed the position with Deep Blue. Deep Blue assessed the sac- 
rifice as sound, but there was no clear win for him. After seeing 
some of the attacking and defensive moves that Deep Blue came 
up with, Garry was glad that he did not make the sacrifice. 
However, he said that he would still have played the sacrifice 
against any other opponent, human or electronic. Deep Blue 
would have played the sacrifice itself, given Garry's position at 
move fifteen. 

After making Deep Blue's move 21,1 left the playing table and 
rushed to the rest room. The match rules specified that the opera- 
tors could leave or return to the table only when Deep Blue was on 
move, or it had just moved. This was somewhat unusual. In 



A Living Mount Everest 1 79 



human vs. human games, either player can leave the table at any 
time. We did not consider the concession too great when the con- 
tract was negotiated, and had agreed to the term. Anyway, by the 
time I rushed back to the game room, Garry was still thinking and 
I was not allowed back to the table. He thought for a long time on 
his twenty-second move, and I waited until he made the move. As 
a result of what happened next, we tried to change the rules for the 
1997 rematch to allow operators to come and go quietly. The rules 
were changed initially, but somehow the old rules were adopted in 
the end. 



8 
7 
6 
5 
4 
3 
2 
1 




abcdefgh 

Position after Carry Kasparov's move 22. f 5. 

When 1 got back to the table, my computer screen had 
entered into a blank "screen saver" mode. I typed in Garry's move 
f5, and Deep Blue crashed. The screen saver mode interpreted the 
'{' in f5 as a wakeup character, and Deep Blue received a '5' instead 
of the full move f5. Joe had been working on the Deep Blue paral- 
lel chess program, and had added some testing commands to the 
program to help him with his work. Unfortunately, '5' was one of 
the testing commands telling workstation number 5 to listen to 
the next command. Needless to say, the command '5' would cause 
an immediate program crash when not executed in the program- 
testing mode. 

Garry was on his way to take a break. The program crash made 
it necessary for me to restart Deep Blue, with our clock ticking 



180 Chapter 10 

away. I informed the arbiter, and started typing as fast as I could. 
Murray and Joe rushed into the game room to find out what went 
wrong. Garry saw the activities, got flustered, and complained to 
the arbiter. Clara, who was in the game room throughout the 
entire match, shouted something in Russian to Garry, and he 
calmed down, perhaps realizing that Deep Blue's clock had been 
ticking. (Someone told me afterwards that Clara had shouted "shut 
up".) After coming back from the crash, Deep Blue quickly 
answered Bxc4. Garry had expected Bxf5, after which he had pre- 
pared to play Nxf7, sacrificing his knight for two pawns. One 
report asserted that Garry claimed the knight sacrifice would have 
led to a winning position for White beyond Deep Blue's search 
horizon. Judging from the faulty analyses given in chess books on 
the match, the knight sacrifice would certainly have won against 
human opposition. Post-game analyses with Deep Blue suggested 
stronger defensive moves for Black than the ones given in the 
books, and some of the attacking moves given in chess books 
appear to lose outright for White. Personally, I believe the sacrifice 
would have led to a draw with best play by both sides as there were 
enough counterbalancing factors. During Garry's visit to IBM 
Research after the match, he also wanted to see Deep Blue's analy- 
sis after his f5. The 1996 Deep Blue preferred Bxc4 slightly over 
Bxf5. The 1997 version of Deep Blue preferred Bxf5. If Nxf7 does 
not work after BxfS as I suspect, then both Bxc4 and Bxf5 are 
playable. 

By the time Deep Blue played Bxc4 in response to f5, most of 
Garry's opening advantage had dissipated. Even so, Deep Blue's 
position looked tenuous. After Garry's move 34, Deep Blue pushed 
its weak c pawn forward one square and suddenly it was threaten- 
ing checkmate. Garry had less than one minute per move left 
before the time control at move 40. For the next few moves, both 
sides made at least one mistake. Garry made his because he was 
short on game time, and Deep Blue's because we were short on pro- 
gram development time. At the time of the match, only one third 
of the positional features recognized by the chess chip were in use 
by the chess program. One key endgame feature for this game was, 
that if all the pawns are on the same half of the chessboard, the 
game is more likely to be drawn. The chess chip could recognize 
this important endgame feature, but the software did not make use 
of this ability. Deep Blue made a move that allowed Garry to 
remove its last queenside pawn. Deep Blue still had the better posi- 



A Living Mount Everest 181 

tion, but he had no problem finding the best way out and sacri- 
ficed his rook for Deep Blue's strong knight. On move SO, a draw 
was agreed. After four games, the match was tied at two points 
each. 

During the next rest day, I saw Frederic Friedel in the hotel 
elevator. He described what happened to Garry, "He went back to 
the hotel room, stripped to his underpants, and stared at the ceil- 
ing for a long time." 

Garry, however, would soon recover. 

A Mount Everest That Crows 

The quest to create a computer that could beat the World Chess 
Champion in a match, was one of the oldest holy grails in com- 
puter science. The quest could be traced back to the 18th century, 
before Charles Babbage's mechanical computers in the 19th centu- 
ry, and well before ENIAC, the first programmable electronic 
computer. In 1949, three years after the creation of ENIAC, Claude 
Shannon, the father of information theory, re-ignited the fire for 
the quest and described how a computer could be made to play 
chess. The holy grail remained hidden for nearly 50 years after 
Claude's seminal work. To a computer scientist, the quest to beat 
the World Chess Champion had taken on the same significance as 
climbing Mount Everest to a mountain climber. 

One slight difference. The World Chess Champion is a living 
being. We are talking about a Mount Everest that grows. 

The first four games of the 1996 match gave computer scien- 
tists a glimpse of the holy grail. We could see the top of Mount 
Everest; the problem was, that this Mount Everest would soon 
grow a few hundred feet taller, right in front of the eyes of the 
world. 

On 15 February, the last rest day of the match, Joel and 
Murray continued their search for an opening line for White that 
could give us another win. By the afternoon, we had found a line 
that appeared to give Deep Blue good chances against Garry's 
Sicilian defense. Still smarting from the debacle in game three, this 
time we started checking the new line during the rest day. The next 
morning, going over Deep Blue's analyses of the possible continu- 
ations, we found and patched a few holes. 

Deep Blue played the same opening move as it did in game 
one and game three, e4. Garry instantly replied eS. I did a double 
take. No Sicilian! We had gotten him to abandon his Sicilian 



182 Chapter 10 

defense! Was he afraid of our treatment of the c3 Sicilian? Calm 
down. This also meant that our opening preparations for the last 
day and half were wasted. Both teams were in new territory now. 
Or in the case of Deep Blue, two-month old territory 10 . Two 
months before the match, Murray spent a few hours putting in 
some alternative book lines, just in case that Garry played some- 
thing else. Very little testing had been done with these alternative 
lines since. 

The game followed with Nf3 and Nf6. Garry wanted to play 
the Petroff defense, also known as the Russian opening. Petroff is 
known to be somewhat drawish. Deep Blue avoided the Petroff and 
went into the Four Knights Scotch opening. The fact that we had 
never tested the opening showed in the way Deep Blue played the 
game. By move 23, Garry had gained a slight advantage; after mak- 
ing it, he offered a draw. 




HAS 



f g h 



Position after Garry Kasparov's move 23. 

I was the operator at the board. The Deep Blue team had 
agreed before the match that a draw offer would be relayed to the 
rest of the team in the back room, and the decision would be made 



'0 Garry had a very odd comment about Deep Blue's immediate book response to 
his opening choice in game five. "How could you know that I would play this?" 
he exclaimed. Garry rarely played this opening line, but what he played was per- 
fectly normal in Grandmaster play. He seemed to believe that we would only 
prepare for the opening lines that he usually played. 



A Living Mount Everest 183 

from there n . No one on the Deep Blue team was expecting a draw 
to be offered at this early stage. The two earlier draws had been 
played until clearly drawn positions were reached. I rang the back 
room and told the gang about the offer. 

Deep Blue assessed the position as slightly better for Garry. I 
was a little nervous about the position on the board. Garry had one 
extra pawn on the king wing, and it was cramping Deep Blue's 
position. Back in the Deep Thought II days, 1 once played through 
a game where one side had such a kingside pawn majority, and 
launched a deadly attack using the extra pawn to great effect. Deep 
Thought II had serious problems understanding the value of the 
kingside pawn majority in attacks against the king. Deep Blue had 
the potential to understand the position on the chessboard, but I 
suspected that the software coding had not been done yet. With 
Garry looking on, I could not voice my concern to the team. The 
rest of the team would have to make the decision. Besides, accept- 
ing a draw this early, when there was still a lot of play left in the 
game, seemed like the wrong thing to do. It was not every day that 
you get to sit across from the World Chess Champion in a serious 
game. No decision came from the back room. Meanwhile, Deep 
Blue had made up its mind on what it wanted to play, unaware of 
the draw offer. Mike Valvo, the arbiter, made a snap ruling insist- 
ing that we either had to accept the draw offer immediately or 
make Deep Blue's intended move. Mike's sudden ruling came as a 
shock, and even today I consider it somewhat questionable. In 
chess games between human players, the draw offer stands until it 
is declined or until a move is made on the chessboard. No move 
had been made on the chessboard, and the match rule explicitly 
stated that the operator could accept the draw offer with or with- 
out consulting Deep Blue. In other words, when the draw offer was 
made, the decision was no longer Deep Blue's, and the fact that 
Deep Blue had made up its mind on the move to play, should not 
have affected how or when the operator made the decision. Mike 
Valvo's ruling forced us to make up our mind immediately, and the 
draw offer was declined. 

What then followed was not pretty, unless you were Garry. 



11 The match rules did not explicitly allow this arrangement. All parties involved, 
including the Kasparov team and the match arbiter, recognized it beforehand as 
the only practical solution, given the operators' relatively weak chess playing 
strength. 



184 Chapter 10 

Deep Blue showed several weaknesses as the game unfolded. Some 
of the weaknesses could be fixed by simply activating the unused 
hardware evaluation terms on the chess chip. Others required 
adding new evaluation terms either in hardware or in software. On 
move 47, twenty-four moves after Kasparov's draw offer, we 
resigned for Deep Blue. At the time the resignation took place, 
Clara Kasparova gave a visible sigh of relief. Garry just needed one 
draw for the last game to win the match, and he would have 
White. 

That night and the next morning, Murray and Joel worked 
hard. To draw the match, we would have to win the last game with 
Black. It was a tall order. Garry lost very few games as White in his 
career. Moreover, our opening system for Black had been prepared 
under the premise of trying to draw as Black. No opening was ever 
prepared to handle the case of having to win as Black. The open- 
ing we played in game four was barely adequate to try to hold the 
draw. We needed something else, something that was riskier but 
held some chance to win. Murray and Joel settled on the idea of 
playing the Semi-Slav defense. They tried to enter as many possi- 
ble variations as possible, and did whatever tests they could in the 
short time available. 

Game six followed a normal line for the Semi-Slav defense. 
Then Garry deliberately transposed his moves to take Deep Blue 
out of its opening book. The fact that we had never let Deep Blue 
Jr play this opening became immediately obvious. Deep Blue was 
supposed to develop its light-squared bishop to the b7 square, but 
after being taken out of the opening book by Garry's transposition, 
it moved the bishop to the d7 square instead. Deep Blue liked to 
put its bishops on open diagonals, diagonals no longer impeded by 
pawns, and it did not understand that sometimes it is acceptable 
to put the bishop on a closed diagonal, as long as the option to 
open the diagonal existed. It was possible to fix the problem to 
some extent in software, but a complete solution would require 
modifying the chess chip. The modifications, implemented as a 
result of this game, to the chess chip to improve the handling of 
the bishops and, similarly, the rooks turned out to be critical in the 
1997 rematch. 

Today, whenever I look at the remainder of the game six, 
especially with the help of the newer version of Deep Blue, I can- 
not help but wonder, "How the hell did the old Deep Blue manage 
to tie the match after four games?" The old Deep Blue made many 



A Living Mount Everest 185 

positional mistakes in the game, and Garry, sometimes cajoling, 
sometimes pressuring, had everything to do with causing the mis- 
takes. The last major positional mistake by Deep Blue was to move 
its bishop to b8 on move thirty, trapping both the bishop on b8 
and the rook on a8. The chess chip was able to detect the bishop 
being trapped. Unfortunately, or perhaps fortunately in the sense 
of shortening our torment, there was very little penalty levied by 
the evaluation software for getting the bishop trapped. After 
Garry's move 43, we had suffered enough and resigned for Deep 
Blue. He had won the match by the score of 4 to 2. 

In the first four games, Deep Blue was treading familiar 
ground and was able to avoid making overt positional mistakes, 
most of the time. In the last two games, Garry forced Deep Blue 
into unknown territory, probed for its weaknesses, exploited them, 
and played Deep Blue as if he were a virtuoso conductor. Mount 
Everest grew a thousand feet. 




Deep Blue 1 chess chip, as used in the 1996 match. 



CHAPTER 11 



Retooling 



Picking up the Pieces 

The drive back from Philadelphia was unusually quiet. Everybody 
was exhausted, both physically and mentally. I was not in the best 
of moods, but I was feeling much better than during the drive to 
the match. Losing is never pleasant, but I knew in my heart that 
we would get another shot at Garry. 

What we had achieved in Philadelphia exceeded the expecta- 
tions of almost everybody, including ourselves. After the first four 
games, we had tied the match with possibly the best chess player 
ever. We could easily have drawn game two if we had more time to 
test the program thoroughly. In game four, Garry had a precarious 
ending that he managed to draw with Deep Blue's help 1 . Deep 
Blue could have been ahead after the first four games. If we had 
been leading then, who knows what would have happened 2 ? The 
publicity generated from the match was tremendous; the match 
news was on the front pages of the New York Times, USA Today, and 
possibly every newspaper in the United States. The 'Tonight' show, 
the 'Late Night with David Lettermari show, and the 'Saturday Night 
Live' show all carried spoofs of the match in their late night TV seg- 
ments. I did not know it at the time, but over one billion 
"impressions" 3 were recorded worldwide during the match. The 



i The 1997 version of Deep Blue would have treated the ending differently, and 
had a decent chance to win. 

2 On the other hand, the last two games of the match did show that there was still 
a gap, and the 1996 Deep Blue, as we suspected, did not really deserve to win. 
Looking back, I am glad that we lost the Philadelphia match. The better player 
won. 



186 



Retooling 187 

equivalent advertising value for the US market alone was worth 
over $100,000,000. We were within striking distance of beating 
Garry and making history. The event could only get bigger. There 
was no way in hell that IBM would not want a rematch. 

I took a day off, and then went back to the office out of iner- 
tia. Murray and Joe took a week off. Each had a newborn baby at 
home, and much catching up to do. Joel's contract with IBM had 
expired, and he returned to the life he knew before his short stint 
on the Deep Blue team. I was realty in no condition to work, and 
was merely doing things that did not require much thinking. 
Before the Philadelphia match, I wrote a software evaluation func- 
tion that used all the evaluation features available on the chess 
chip. This was written to do experiments on automatic adjustment 
of the evaluation weights, also known as evaluation tuning, for the 
hardware evaluation features. The program measured how well the 
program's moves, obtained from very shallow searches, matched 
Grandmasters' choices in positions from their games. The basic 
assumption was that the greater the matching rate, the better the 
evaluation function, and the stronger the chess program. Running 
the experiments did not require a lot of thinking, and since there 
was ample machine time available after the Philadelphia match, I 
spent all my days chewing up machine cycles to run the experi- 
ments. 

These experiments on evaluation tuning uncovered some sur- 
prising facts. In particular, the weights assigned to "rooks on open 
files" by human Grandmasters, appeared to be much lower than 
the ones traditionally used in chess programs. I knew about this 
before the Philadelphia match, but I thought the result an artifact 
of some software bugs that I had introduced somewhere. Closer 
examination after the match indicated that the result was real. It 
was not that human Grandmasters attached lower weights to 
"rooks on open files", but that they were far more discriminating 
about when "rooks on open files" were important. In other words, 
there were far more contexts used by the human Grandmasters, 
either consciously or unconsciously, than by typical chess pro- 
grams. After this discovery, I went through the entire set of 



i "Impression" is a term used in advertising. Just hearing, seeing or reading the 
advertising message does not count as an impression. Only after the message is 
registered in the mind of the potential customer does the "impression" count. 
Each impression usually requires several exposures to the message. 



188 Chapter 11 

hardware evaluation features, and located all the features that had 
low evaluation weights according to the tuning program. All these 
features might be in need of improvement. 

Murray, Joe, and I had a meeting after they got back from 
their vacations. We examined what went wrong in the match, and 
reached three conclusions. The first was that Deep Blue's search 
speed appeared to be adequate. Garry was outplaying Deep Blue 
positionally, but not tactically. This brought us to our second con- 
clusion, which was that Deep Blue was grossly inadequate in the 
amount of chess knowledge that it had, compared to the World 
Champion. The chess chip was too new for us to learn how to use 
the hardware evaluation features effectively in the short time avail- 
able. During the Philadelphia match, only about one-third of the 
hardware evaluation features were really in use. There was a lot of 
room for improvement by creative use of the existing chess hard- 
ware. The third conclusion was that the team, the humans, had to 
be better prepared. We were not prepared when Garry made his 
draw offer in game five. For game six, we did not have an opening 
ready to try to win as Black. Psychologically and procedurally, we 
were not ready for the big match, period. The team had gone 
through a lot together, but none of us really had any big match 
experience. 

We had no idea when we would get another shot at Garry. Joe 
had some doubt whether we would get another chance at all, but 
after talking with Murray and me, he was convinced and relieved 
that we would probably have another go. The question was what 
we could do to increase our odds next time. We could improve the 
search speed of the program, but that probably would not help 
much. The software evaluation function could certainly be 
improved. We could also improve the chess hardware, but that 
would be possible only if there was enough time before the 
rematch, and there was enough money allocated to make a new 
chip release. We would need to do whatever we could to convince 
the management to schedule the rematch as late as possible, and 
to allocate an adequate budget to allow for the release of a second- 
generation chess chip. Murray and Joe were not too thrilled about 
the prospect of going through the gyration of creating a new chess 
chip, and all the new software to go with it. But I persuaded them 
to at least make sure that we had adequate time and funding just 
in case. We all agreed on what the last set of tasks should be. We 
needed to build tools for match preparation. Garry won the 



Retooling 189 

Philadelphia match in part because he could adapt faster than 
Deep Blue. It was difficult to make Deep Blue as adaptive as a 
human being, but we could help by working as a well-oiled team. 
In Indy 500 car races, after going around the laps several times the 
cars need to get their tires replaced at the pit stop. The pit-stop 
mechanics, by working as a team, and with the proper tools, can 
replace the tires in less than eleven seconds. We did not need to be 
as fast as pit-stop mechanics, but we could adopt the same 
approach. We needed to create sets of tools for everyone involved: 
a set of tools for the programmers (us), and a different set for the 
Grandmasters. We also needed to make sure that all members of 
the team knew what their responsibilities were. 

Before the new tools were created, however, I got into a debate 
with Murray and Joe on whether there should be a new chip, and 
then on what should go into the new chip. This debate lasted sev- 
eral months. 

Another Six Months 

For some unknown reason, I seem to be able to concentrate well 
after a long overseas trip. In 1985, after my trip to Taiwan, I was 
able to pour in the energy needed to finish the move generator 
chip, in six months. In 1996, about a month after the Philadelphia 
match, I was asked to give a talk about Deep Blue at the IBM Tokyo 
Research Lab. IBM Research was having its own fiftieth anniversary 
celebration, and the Tokyo Lab was hosting a one-day conference 
on parallel processing as part of the celebration. My talk was one 
of the sessions in the conference. For the next six months after this 
Tokyo trip, I worked harder than I ever had in my entire life, hard- 
er even than when I designed my first chess chip. 

In 1985, there was no question over whether the chess move 
generator would be fabricated, as the MOSIS service was effective- 
ly free. In 1996, I began work on the new version of the Deep Blue 
chess chip without knowing whether there would be money to 
build it. Besides working on the new chip, I also had to wear the 
hat of an advocate, first persuading the rest of the team, and then 
our manager, that the new chip was desirable and necessary. 

Given all the uncertainty, I took a staged approach, both in 
the chip design and in my attempt to persuade everyone. New 
functions were added to the chess chip in stages, always maintain- 
ing a new "working" version in the process. This meant I could 
stop the new design at any point, and have something that was 



190 Chapter 11 

ready for fabrication on very short notice. The chip design process 
was relatively straightforward, I just had to put in the work; the 
task of being an advocate was much harder. 

Originally, money had been set aside only for fabricating 
more of the same chess chips that were used in the 1996 match. It 
took some doing, but I managed to convince the rest of the team 
that we should at least add a repetition detector to the chess chip 
before going through another chip run. Murray, Joe and I, had 
observed Deep Blue Jr playing strange moves that appeared to 
result from the "horizon effect" related to the lack of repetition 
detection in the chess hardware. Deep Bluejr would sometimes try 
to repeat a position that it liked by pushing the position into the 
part of the search tree beyond its "repetition detection horizon", 
which was the "horizon" for the software search tree that did have 
repetition detection. While this behavior had a definitely detri- 
mental effect, neither Murray nor Joe was willing to endorse the 
addition without clear evidence that it was necessary. The disasters 
that we had in Hong Kong turned out to be the persuasion that I 
needed. The draw given up to WChess could have been avoided. 
Many other things went wrong simultaneously in our loss to Fritz, 
but with a hardware repetition detector, the machine would have 
avoided the problem line in the first place. It was conceivable that 
Deep Thought II might have won the Hong Kong event with a per- 
fect score; in other words, its performance would have been more 
than 200 rating points higher. This final argument got both 
Murray's and Joe's attention, and they agreed to endorse the idea. 
This took place before my visit to Tokyo. 

Some time before the Philadelphia match, I realized that there 
was a better way to implement the evaluation function that would 
reduce the chip area significantly. During the flights to and from 
Tokyo, it dawned on me that this implementation also made it 
possible for the evaluation function to be far more comprehensive. 
The problem was that this would entail a wholesale change to 
roughly two-thirds of the chess chip. The change was highly desir- 
able, but it was also very risky. I believed that this change had to 
be done if we wanted to beat Garry, and I really wanted to win. 

We had a team meeting with C J after my trip to Tokyo; 
Murray and Joe helped me persuade him that the detector had to 
be included in the new chip. But all three balked at the idea of cre- 
ating a brand-new evaluation function. In the end, it was agreed 
that I would design two versions of the chess chip, both with the 



Retooling 191 

hardware repetition detector, but one with the old evaluation 
function, and one with the new. My first priority was to create the 
version with the old evaluation function. I could spend time on the 
version with the new evaluation function, but I would not be get- 
ting any help from Murray or Joe — they had two sets of match 
preparation tools to create. Of the two new chips, only one would 
be fabricated. To get the new evaluation function, 1 would have to 
complete the design of the repetition detector, and the new evalu- 
ation function, as well as having the entire chip thoroughly 
verified, all by myself. 1 had four months until we would decide- 
which version to use. 1 agreed to the extra workload as it provided 
another chance to create a chess machine that was strong enough 
to beat the World Chess Champion, and I was going to do every- 
thing within my power to make that happen. 

The first month was relatively easy, and mostly spent design- 
ing and verifying the repetition detector. By the end of the first 
month, I had the first new version ready. There were three months 
left for the even newer version that I really wanted. Based on the 
experience from the first Deep Blue chess chip, Murray and Joe sur- 
mised that it would have taken at least a solid year even with their 
help. There were several reasons why I believed that I could pull it 
off. First, there were quite a number of software tools left over from 
the design of the first Deep Blue chess chip, and I could leverage 
the old work to save some time. Second, I was rested and I could 
pour in longer hours if necessary. Third, being the sole designer, I 
had the full picture of what was in the design, and I would have 
much shorter cycles of finding problems and fixing them. Finally, 
above all, I wanted to win. 

The second month was spent redesigning the new evaluation 
function. Since it was a wholesale change, 1 could not tell whether 
anything was working until the new design was partially complet- 
ed, about two weeks into the second month. 1 was thrilled that the 
new circuit passed my initial test — the new design worked in prin- 
ciple. Now came the hard part. A software model had to be created, 
and the design and the software model had to match each other. 
The software model had to be created based on the design intent, 
not based on what was in the design, to avoid perpetuating any 
design error. This software model was almost completely new, as 
the new evaluation function was quite different from the old one. 
Pieces from the software model for the old chip were used, but 
mainly as building blocks. For the next month and half, until the 



192 Chapter 11 

end of the third month, I was working close to 100 hours per week. 
There were no weekends. Every night when I drove home from the 
lab late at night, 1 would tell myself repeatedly, "I want to win." 
After I got home, I would check my circuit simulation to make sure 
it was still running. If I woke up during the night, I would check 
the simulation again, and if necessary, fix my software model or 
whatever before restarting the simulation. 

By the end of the third month, 1 had the new evaluation func- 
tion reasonably verified. Of course, this would have not been 
possible without the software tools left over from the old chip. I 
was starting to get ambitious. The software model, and the hard- 
ware were matching quite well, and I no longer had to wake up in 
the middle of the night to check things. Since I was ahead of 
schedule, I figured that I might as well look for whatever other 
improvements I could make. Before the first Deep Blue chess chip 
was fabricated, my original design had a complicated chess move 
generator that could "prune" away provably useless chess moves, 
which theoretically could effectively speed up the chess chip by 
three to five times. The plan for the complicated move generator 
was dropped in the first Deep Blue chess chip, as the correctness of 
the computation involved was quite difficult to verify. There was 
not enough time to do the original move generator that I planned 
earlier. It was far too complicated. But I could do something inter- 
mediate in complexity if I was willing to drop the "correctness" 
requirements on how I pruned away the chess moves. I found a 
pruning scheme that I was happy with, went ahead and imple- 
mented the new move generator, without telling either the rest of 
the team or our manager. To be on the safe side, I added to the new 
move generator a mode that allowed it to behave the same way as 
the old move generator. Eventually, after it was completed and test- 
ed, I did inform Murray and Joe. They were not thrilled. 

The new move generator, along with some improvements in 
other areas, was designed and verified in the fourth month. As the 
end of the fourth month drew near, it became clear to Murray and 
Joe that my new design was ready and we made the conscious deci- 
sion to tailor all the new software development to the new design. 
The new evaluation function offered so many powerful new fea- 
tures that both Murray and Joe decided that making the switch 
would be worthwhile. Even if there were some minor problems 
with the new chip, the advantages provided by it were over- 
whelming. 



Retooling 193 

It turned out that there were two more months to go before 
the design had to be delivered to VLSI Technology. 1 used the time 
to make small incremental changes. 

Meanwhile, after we made the decision to use the new evalu- 
ation function, Murray took its software model, added a graphical 
interface, and created a powerful analysis tool. This new piece of 
software became an important tool during the program develop- 
ment, and during the rematch itself. Joel rejoined the group at 
about this time, and was introduced to the tool. Quite likely, the 
most important new evaluation features on the new chess chip 
were added as a result of our interactions with him. 

One day, while I was going over a new evaluation feature that 
I just added, a very excited Joe Hoane came into my office. He said, 
"There is something that you should see", and took me to Joel's 
office. Murray was already there. Apparently, Joel had been talking 
with Murray and Joe about some problems that Deep Blue Jr had 
in a game Joel had just played against it. He had seen similar prob- 
lems before, but could not pinpoint the exact source until he got 
the new analysis tool for the evaluation function. 

Deep Blue Jr, as well as most of the other modern chess pro- 
grams, understands the importance of putting rooks on open files, 
files that are no longer impeded by pawns, so that the rooks can 
exert the greatest influence. The problem was that, at the time, 
along with any other chess programs that I knew of, Deep Blue Jr 
did not know that rooks could be profitably placed on files where it 
had the option of "opening up" the files. That is, the option of 
removing the pawns impeding the rooks, usually by making a pawn 
capture. In the game that Joel just played, Deep Blue Jr could have 
piled up its rooks on such a potentially open file and exerted long 
term pressure indirectly. Instead, Deep Blue Jr opened up the file 
prematurely and then placed its rooks on the resultant open file. 
But this action gave Joel the chance to challenge the open file with 
his own rooks, which then were traded off along the open file. With 
all the rooks gone, Deep Blue Jr's positional advantage dissipated 
into thin air. Joel suggested that the right way to play the position 
was to just pile up the rooks on the potentially open file, and wait for 
the most opportune moment before opening it. Similarly, the same 
concept also shows up in the relationship between the bishops and 
the diagonals. Game six in the first match was a prime example of 
Deep Blue's incorrect treatment of bishops on diagonals. After some 
discussion, we came up with rules for recognizing the relevant fea- 



194 Chapter 1 1 

tures. That very afternoon, I put in and tested the new features 
which was a relatively easy modification. This new feature showed 
up in the critical game two of the 1997 rematch when Deep Blue 
pinned down Garry's heavy pieces to his back rank as a result of a 
potentially open file. Without the new feature, it was very unlikely 
that Deep Blue would have won. Not bad for an afternoon's work. 
Other major new evaluation features were also added during 
the last two months of chip design. The first Deep Blue chess chip 
had a fairly comprehensive evaluation for king safety, but the 1996 
match provided ample evidence that still more work was needed. 
A massive redesign of the king safety evaluation logic took place 
during the last two months of chip design, with the help of the 
evaluation function analysis tool. This redesign positively influ- 
enced Deep Blue's play in games five and six of the rematch. 
Undoubtedly, it also altered the course of the other games at one 
point or another, for better or for worse. 

The Phantom Queen 

The netlist for the new chess chip was sent out in September 1997 
to a design firm contracted by VLSI Technology to do the physical 
design. Fabrication for the new chips started in late December 
1996. The printed circuit boards for the new chess chips were 
designed in parallel, and became available before the new chips 
were shipped back in February 1997, less than three months before 
the rematch. After long negotiations with Garry, the rematch had 
been scheduled to take place in May. The amount of time left for 
preparation for the rematch was only slightly better than the time 
left for the first match. However, the software was in much better 
shape this time. We already had working software based on the old 
chips, and we had the new software evaluation function in pretty 
good shape before we ever laid eyes on the new chips. 

The initial integration of the new chess hardware with the 
new software went smoothly, but then we started to hit snags. Joe, 
who was doing the bulk of work for the software integration, did 
the integration methodically. He started with the common features 
of the old and the new software, then added new features one at a 
time. He first ported the old software to the new chess hardware, 
and it went fine. He then simplified the software interface to the 
chess hardware, as the new chess hardware allowed a simpler inter- 
face. This also went well. The next step Joe took was to activate the 
hardware repetition detector where he soon ran into problems. 



Retooling 195 

The design software that I used to simulate the new chess chip 
was somewhat limited. Ideally, we would have liked to be able to run 
the chess software on top of the simulated chess chip, so that any 
problems with integrating the chess software with the chip could be 
discovered before the chips were fabricated. There were two obsta- 
cles that faced us. First, there was no easy way to do it. The design 
software did not have a simple mechanism that allowed us to drive 
the simulated chess chip directly from the chess software. Second, 
the simulated chess chip was over a million times slower than the 
real chess chip, and one minute of running time on the real chip 
would have cost us over two years of simulation time. There was only 
just slightly over one year of time between the 1996 and the 1997 
match. Usually, it would take at least several hours of running time 
on the real chip to uncover all the problems associated with the sys- 
tem integration. So the new chess chip was never really verified to 
my complete satisfaction before it was submitted for fabrication. 
More advanced design software was already available, but there was 
no time to convert the design to run with the newer design software. 
Besides, it cost big bucks to get the newer design software. 

The hardware repetition detector was thoroughly tested, or so 
I thought. 

In 1992, when I transferred the control of the chess software 
to Joe, I wrote the chess program in a certain way. All my testing 
of the hardware repetition detector was based on how I interfaced 
the Deep Thought II software with the chess hardware. Joe had 
changed and improved the software substantially in the interven- 
ing four years. By early 1996, his new program for the search part 
alone had grown to well over 100,000 lines of code, or about 20 
times the size of the code that I gave him in 1992. Needless to say, 
Joe accessed the chess hardware in a more complicated fashion, 
and used the full set of all possible operations. I simulated and test- 
ed the hardware repetition detector based on how 1 wrote the chess 
program, but with Joe's code, it did not work properly. Some of the 
chip operations that were not tested in my simulation, but were 
used by Joe in his code, had the side effect of giving false data to 
the hardware repetition detector. 

The problem was not fatal, and the solution was relatively 
easy, as long as you didn't have to write the code. He modified the 
chess program to use the smaller set of chip operations that worked 
properly with the hardware repetition detector. This cost us about 
a week but Joe got it working in the end. 



196 Chapter 1 1 

The next problem that he uncovered, however, had no pleas- 
ant solution. One of the big questions about the new chip was how 
well the hardware search pruning really worked. After Joe enabled 
the hardware search pruning, the early results were fairly good. 
The pruning appeared to speed up the search by a factor of 5 to 10, 
more or less in line with my estimate. Then Joe found something 
peculiar. In one of the positions that he was testing, an extra White 
queen showed up on the al square (lower left corner) of the chip's 
internal chessboard after the search was completed. The problem 
was repeatable no matter which chip was used. It was almost cer- 
tain that I had introduced a logic design error somewhere. The 
immediate questions were what and where? 

Joe located a position where the phantom queen showed up 
in a three-ply search. It took about one day of simulation to repli- 
cate the error. My least favorite chess rule came back to haunt me. 
The phantom queen only showed up when en passant was possible 
in the position. So it was en passant yet again. Under certain rare 
conditions, either a White queen on al, or a Black queen on a8 
could materialize out of thin air. The phantom queens, however, 
would not show up if we disabled the hardware search pruning. 

We had three ways to solve the phantom queen problem. 

The simplest was to just disable the hardware search pruning, 
giving up the five- to ten-fold speedup from the pruning. This was 
adequate as a short-term solution to get something running imme- 
diately. 

The second approach, which I preferred, was not feasible 
time-wise. It was a hardware solution. The new chess chip was 
designed to be controllable externally with an FPGA (Field 
Programmable Gate Array), costing around $20. The external FPGA 
could monitor the hardware search and remove the phantom 
queens as they showed up. This approach would also give us the 
option of adding additional evaluation features inside the FPGA, 
without going through the costly redesign of the chess chip. 
Another advantage of the hardware approach was that the software 
remained essentially the same. There were slightly over two 
months left before the rematch. This hardware solution required a 
new batch of printed circuit boards, and our regular printed circuit 
board supplier said that at best they could supply the boards about 
two weeks before the rematch. Two weeks was too tight even for 
my taste. We might have been able to find a new supplier who 
could supply the boards faster, but it was too risky a proposition. 



Retooling 197 

The third approach required massive software changes, and I 
was of the opinion that we were better off sticking with the sim- 
plest solution of the first approach. I was not comfortable with 
making such changes right before the match. This approach was 
based on the observation that whenever the phantom queen 
appeared after the hardware search, the positional evaluation 
almost always changed slightly. The proposal was to enable the 
hardware search pruning before the search as the default, and 
when a change in evaluation was observed after the default search, 
remove the phantom queen(s), disable the hardware pruning, and 
re-search the position. Normally, less than one in ten positions 
given to the chess chips needed to be re-searched. The effective 
speed was thus about thirty to one-hundred percent slower than 
the design speed, but much faster than the speed provided by the 
simplest software fix. There were two correctness problems with 
this approach. First, there was no guarantee that a positional eval- 
uation change would be observed when a phantom queen did 
show up after hardware search. Second, the phantom queen, given 
the right conditions, could show up during the search, but then 
get overwritten before completing the search, and thus cause a 
false search result yet without any indication of the fault. In other 
words, the search could not be completely trusted because of the 
phantom queen problem, if the third approach was adopted. It was 
an intellectually interesting solution, which was probably why Joe 
decided to spend about a month getting it to work. 

There were no other hardware errors uncovered after the 
phantom queen problem. The problem ended up costing us about 
one month of Joe's time, and probably made the software less sta- 
ble as a result. The last known software bug caused by the phantom 
queen problem was not fixed until two weeks before the 1997 
rematch. In the end, everything worked out all right, but things 
did look bad for a while. 

Quality before Quantity 

During the 1996 match, Garry congratulated the team for succeed- 
ing in "converting quantity into quality". What he did not realize 
at the time was that Deep Blue never had just quantity. In particu- 
lar, in the drawn game three, some of the new evaluation features 
in the chess chip came prominently into play. Some of the "inhu- 
man" moves were not the result of deep searches at all, but of Deep 
Blue evaluating Garry's rook getting close to being trapped. 



198 Chapter 11 

For the 1997 rematch, the main emphasis of our match prepa- 
ration was to improve the underlying quality of Deep Blue's chess 
knowledge. This time, we wanted the quality to go in before the 
quantity. 

Creating the new chess chip was just the first step in our quest 
for quality. 

Going into the 1996 match, Deep Blue had only existed for 
two weeks. It was a baby. It was a very powerful and very talented 
baby; otherwise it could not have possibly beaten Garry in its very 
first game. But the result of the 1996 match also made it painfully 
clear that the baby needed to be taught the tricks of the trade. By 
September 1996, it was certain that the rematch would take place, 
and Joel Benjamin was invited back to IBM Research to be the 
headmaster of a chess school with only a single pupil, namely, 
Deep Blue. (In reality, most of the time, the pupil was actually 
Deep Blue Jr, since we did not have full access to the IBM RS6000 
SP supercomputer that Deep Blue ran on.) 

A typical school day started with Joel playing "take-back" 
chess with Deep Blue Jr; sometimes this continued in the early 
afternoon. Deep Blue Jr was probably already of Super 
Grandmaster (2600+ international rating) strength even in the 
early days, so Joel allowed himself to take back his own moves 
when the position started to look bad for him. By going back 
repeatedly to positions where he still had chances, he found lots of 
positions that caused Deep Blue Jr serious problems. After the 
"take-back" chess sessions, he would tell us about the mistakes 
made by Deep Blue Jr. We would then either make changes to Deep 
Blue Jr or, in some circumstances, make sure that the problems 
were dealt with in the new chess chip. The next day, the whole 
process started again, usually with some surprises waiting for Joel. 
The newly improved Deep Blue Jr would frequently play moves 
that it would not have played the previous day. 

Most of the daily changes were in its evaluation function. 

Up to this point in time, the Deep Blue team acted as a troi- 
ka with each team member responsible for a disjointed area of 
expertise. Murray was responsible for the evaluation function and 
the opening book, Joe had all the rest of the main program, 
including the very complicated parallel search code, and I was 
responsible for the hardware and the test programs. The team was 
set up so that no one person had the role of the team leader but 
all were equal. Usually, everyone kept to his own area of expertise. 



Retooling 199 

On two occasions, however, we joined in to help one of us who 
was in need. The first time was when the first Deep Blue chess 
chip was near completion. Murray and Joe chipped in to help me. 
Now, with the constant stream of evaluation problems uncovered 
by Joel, and a new evaluation function to complete, Murray could 
not handle the workload all by himself. 1 was still spending most 
of my time on the new chess chip, so Joe provided the much- 
needed help. Murray and Joe complemented each other extremely 
well. 

Murray and Joe had quite different programming styles. 
Murray liked to think things through before he wrote a program. 
This meant that he tended to make steady but slower progress. This 
fact, in conjunction with the short time that we had, was why only 
about one-third of the hardware evaluation features were in use at 
the time of the 1996 match. Joe was more an experimentalist. He 
would write code very fast, try it out and see how it worked. With 
all the new evaluation features suddenly available, Murray was 
overwhelmed with riches. Joe took it upon himself to work on the 
evaluation function codes for the new features, and quickly came 
up with new stuff on top of what Murray had already written. Joe 
was a much weaker chess player than Murray, but the evaluation 
weighting he put in the code turned out to be pretty good. 
Murray's work on the evaluation function became more fine- 
tuning than coding. We ended up getting the combination of both 
Joe's speed and Murray's thoughtful care. 

When the new chess chips came back from VLSI Technology, 
we had a fairly good evaluation function in place. Not all the eval- 
uation features were used to their full potential even up to the end 
of the rematch itself, but the function was already in much better 
shape than what we had at the time of the 1996 match. Murray 
played the first few games with the new software, using the new 
chess chip, against two of the top commercial PC programs. At the 
time, the new software used only one chess chip, running at a 
slowed down clock rate. Furthermore, because of the phantom 
queen problem, the hardware search pruning was disabled, slow- 
ing down the new software by another factor of five to ten. The 
"Pico" Deep Blue Jr was thus running at effectively only 100,000 to 
300,000 positions per second, or roughly the same search speed as 
the fastest commercial chess program running on, say, a high end 
Pentium Pro-based personal computer. The biggest difference 
between the programs would be in the evaluation function; that is, 



200 Chapter 11 

the chess knowledge encapsulated in the programs or, in the case 
of "Pico" Deep Blue Jr, in the chess hardware. 

Murray played the first few games with the "Pico" Deep Blue 
Jr without telling anybody. One day, he mentioned that he had 
played a few games against the top commercial chess programs, 
and asked me to guess the outcome. Given that Pico did not real- 
ly have an edge in search speed, and the new evaluation function 
was still not tuned, I answered, "About 5#-50." Then Murray 
dropped the bombshell. Pico had won all the games. I wanted to 
see the evidence. Murray showed me fragments of the games from 
memory. The reason for the domination by Pico became obvious. 
In our quest to improve Deep Blue's chess knowledge, we had cre- 
ated something way beyond what had been done in commercial 
chess programs. In each game fragment that I saw, the relative 
chess knowledge deficiency of the commercial programs became 
crystal clear, and I could tell precisely what hardware evaluation 
features were at play in each game. 

Murray played a few more games afterwards. Pico kept on 
winning. On the eleventh game, Murray decided to give a two to 
one time odds to one of the commercial programs. Pico lost the 
game. The other program's relative tactical strength became too 
much for Pico's better evaluation function to compensate. Shortly 
thereafter, Joe completed the first version of the new Deep Blue Jr 
software that used multiple chess chips. We decided that it was the 
time for the new Deep Blue Jr to attend Grandmaster Joel's 
Excellent Little Chess School. At the time when Murray abandoned 
the games against the commercial chess programs, the 10-# record 
by Pico (excluding the game with time odds) indicated that, in 
computer vs. computer play, Pico was at least 200 rating points 
stronger than the best commercial chess programs, with roughly 
ninety-five percent certainty. 

The Communication Department in IBM Research started the 
media kickoff for the rematch in early February, at roughly the 
same time we received the new chess chips. One of the kickoff 
events was the first game between the old Deep Blue Jr and the 
new Deep Blue Jr in late February. 

The reporters witnessing the one-game informal private 
match, between the old and the new, were from the ABC 'Nightline' 
show and Time magazine. We were expecting the new version to 
win, but nothing could be taken for granted. After all, we only just 
had the new program working. The game started roughly equal, 



Retooling 201 

but then the new Deep Blue Jr initiated what looked like a prom- 
ising attack. The old Deep Blue Jr was unconcerned, and was in fact 
quite optimistic about its own chances. The new Deep Blue Jr was 
also getting optimistic. Both could not be right. Then, suddenly, 
the new Deep Blue Jr saw the win of material. The old version still 
saw the position as about equal but was definitely wrong. It played 
a move unexpected by the new version. Now, it was all over; the 
new version declared that it would win massive material. After the 
move of the new program, the old program saw the doom as well. 
The game lasted only about twenty moves. Everyone was amazed. 
I would not have believed that any chess playing entity could do 
this to the old Deep Blue Jr. All the chip work on improving the 
king safety evaluation, and all the new software work to exploit the 
hardware evaluation features, had paid off handsomely. The 
reporter from Time commented that the moves played by the new 
Deep Blue Jr had a distinctly human-like quality to them. Garry 
would be playing a vastly different program from the one he 
played in 1996. 

So the new program was dominating other computer pro- 
grams, including the old version of Deep Blue. It was a good sign, 
but one question remained to be answered. How strong would the 
new Deep Blue be against human Grandmasters? 

Grandmaster Miguel Illescas was an up-and-coming chess star 
from Spain. In the last ten years or so Spain has become a major 
chess country. The strongest chess tournament in the world took 
place annually in Linares, Spain, with Miguel being one of the fre- 
quent participants. As a result, he became quite familiar with the 
top chess players. We had our first contact with Miguel in 1995 
when Deep Thought II lost a two-game match to him. We invited 
him to help with Deep Blue's opening preparations for the 1996 
match, and he did it remotely from Spain. For the rematch, Miguel 
had a two-week time slot, in March 1997, in which he could trav- 
el to the United States, and we took the opportunity and invited 
him over. 

Miguel spent the two weeks working with Joel on the opening 
preparation, but every once in a while he would play a few casual 
games with Deep Blue Jr. He lost most of the games. This was 
another good sign but casual games cannot be taken too seriously. 

Joel and Miguel worked very well together. Joel had intimate 
knowledge of how Deep Blue behaved, while Miguel was in the top 
chess echelon. When Joel had some difficult opening problems, 



202 Chapter 11 

Miguel sometimes would come up with what the top players had 
discussed among themselves about the very same problems. When 
Miguel had questions about Deep Blue, Joel would tell him the 
answers in the common chess language among the Grandmasters. 

After Miguel's visit, Grandmasters Nick DeFirmian and John 
Fedorowicz were invited to help with specific opening systems that 
Deep Blue might use during the match. We kept these meetings a 
closely guarded secret until the start of the rematch to avoid signal- 
ing Garry which openings Deep Blue might play, giving him a 
dangerous edge. There was no point in preparing for over a year, 
painstakingly improving the program, only to get slaughtered by 
Garry's renowned home opening preparations. We were going into 
the match to play Garry, not his home preparations. Just like Miguel, 
Nick and John also played casual games against Deep Blue Jr and lost 
most of them. Again, a good sign. But to be certain, we would need 
to play against other Grandmasters in serious training matches. 

We scheduled two training matches for Deep Blue Jr in April 
1996. The first was against Grandmaster Larry Christiansen, and 
the second against Grandmaster Michael Rohde both of whom had 
extensive experience playing against computers. 

John and Nick were both looking forward to seeing the train- 
ing matches in person. In John's words, "1 would like to see them 
suffer." Larry and Michael were close friends of his. To avoid any 
possible leaks of their involvement before the rematch, it was 
decided that they would work at home. 

Larry played the first of the two two-game training matches. 
In the first game, Deep Blue Jr set up a promising position, and 
then suddenly gave up a pawn for no apparent chess reason. It 
turned out to be a software bug causing it to play a random move. 
Shades of ChipTest's first outing where ChipTest played the first 
move from the move generator, whether good or bad. The Deep 
Blue Jr bug was related to how time control was handled. The bug 
had been fairly recently introduced and Joe had been trying to 
track it down for a while. He thought that it was fixed, but appar- 
ently it was still there. The position looked hopeless, but somehow 
Deep Blue Jr whipped up some tactical magic and it was a repeti- 
tion draw. Joe located the bug and fixed it before the game on the 
second day. Deep Blue Jr won the second game, and won the 
match by the score of 1.5 to 0.5. No bug showed up in the match 
against Michael, which again was won by Deep Blue Jr by the score 
of 1.5 to 0.5. 



Retooling 203 

The performance of Deep Blue Jr against the two 
Grandmasters in games played under regulation time control was 
in the high 2700s, or within about 50 rating points of Garry 
Kasparov. There were too few games to take the performance seri- 
ously, but Deep Blue itself would be five to ten times faster than 
Deep Blue Jr, and possibly 100 to 200 rating points stronger. If 
Deep Blue was not yet stronger than Garry, it was well within strik- 
ing distance. 



The Deep Blue 2 chip used in the 1997 final match. 



CHAPTER 12 



The Holy Grail 



The Chinks in the Armor 

Leading up to the 1997 rematch, we sensed a quite different 
atmosphere several months before the event itself. We felt good 
about our chances. Garry Kasparov was strong, but he was not 
invincible, and Deep Blue was closing in. We were making fast and 
steady progress. Our own assessment right before the rematch was 
that, if Deep Blue was not already stronger than Garry, it was close 
enough and he would not be able to win by more than one game. 
On the other hand, he was sufficiently strong that we seriously 
doubted that we could win the match by more than one game. 
There was a fairly good chance that the match would be drawn as 
well. From our point of view, this would be a perfectly good result. 
If we had, say, six more months to prepare, and if we could main- 
tain our rate of improvement, it was conceivable that Deep Blue 
would be able to win the match by more than one game. But, of 
course, we did not have six more months. Whatever we came up 
with would have to do, and it looked as though the new Deep Blue 
might be good enough. 

Publicly, we kept quiet until about three months before the 
rematch. The new chess chips were kept under wraps, and since we 
were not absolutely sure that they would work out, their existence 
was not publicly known until the match was under way. We had a 
contingency plan to revert to the old chess chips if we discovered 
something terribly wrong before the rematch. 

Garry, meanwhile, was apparently getting nervous. In late 
February 1997, Michael Krantz, the same Time reporter who had 
witnessed the one-game match between the two versions of Deep 
Blue Jr, interviewed him. Michael mentioned to Garry that he had 

204 



The Holy Grail 205 

seen the game and been quite impressed with the human-like qual- 
ity of the moves played by the new program. Garry became quite 
agitated. First, he asked Michael for the game score. When he 
replied that he did not have it, Garry asked whether he remem- 
bered what opening was played. Garry was quite disappointed 
when he found out that Michael did not have any idea what the 
opening was called '. What Garry did not realize at the time was 
that ABC Nightline had a complete video tape of the game, and in 
fact, had broadcast segments of the game earlier on air. Did Garry 
ever find out about the ABC Nightline tape? I don't know. But if he 
did, he would have known that he indeed had something to worry 
about. 

Shortly after Garry had his Time interview, we received an e- 
mail from Frederic Friedel asking if he could visit IBM Research 
with a reporter from Germany. We had a friendly relationship with 
Frederic, but we were concerned he might have the side mission of 
probing for information relevant to Garry's match preparation. We 
would be very polite but alert, to avoid giving Garry an edge. 
Frederic visited the lab in early March 1997, and both the visit, and 
the interview went fine. At the end of his visit, Frederic made an 
unusual comment, "Garry would win easily if he should play anti- 
computer chess. But he will never do it." In anti-computer chess, 
the human player deliberately plays sub-optimal moves to get to 
positions that they think are difficult for the computer to handle. 
In this sense, Garry had already played anti-computer chess in the 
1996 match. We did not give Frederic's comment much thought. 
We had prepared extensively for the possibility of playing against 
anti-computer chess already. We had no idea how far Garry would 
actually go in the coming rematch. One of Frederic's questions 
was, not surprisingly, whether we thought Deep Blue could win 
the match. He did not look too pleased when we gave him the 
straight answer that Deep Blue had a very real chance of winning. 
I did not tell him that I believed Deep Blue would have a better 
than fifty percent chance of at least drawing the match. 

As the rematch drew near, Garry started making various 
requests through his agents either outside of, or in conflict with, 
what was in the already signed match contract. Some of Garry's 
requests were quite reasonable, but others were less so. 

1 For the curious, the opening line played in that game was the Tarrasch variation 
of the Queen's Gambit opening. 



206 Chapter 12 

One request was to use a new chess clock, based on Garry's 
specifications and custom designed by Audemars Piguet, a Swiss 
watchmaker that Garry endorsed in advertisements. Another one, 
related to the new chess clock, was to use an entirely new and 
irregular time control that Garry had recently dreamt up. The 
match contract already specified that the regulation time control 
was to be used, just as in the 1995 Kasparov vs. Anand World 
Championship Match. In the end, Garry's new clock was adopted, 
but not his new time control. This new time control was exceed- 
ingly complicated, and even the new clock did not work with it. 
More seriously, even if Garry's clock could handle his new time 
control, the Deep Blue team had neither the time nor the energy 
to modify the program to deal with his late request. 

Garry was not handling the request himself, so he was not 
wasting his time or energy on it. But if we had agreed to it, we 
would have had to spend valuable time, which we did not really 
have, to code the new time control and do extensive testing which 
could have cost us weeks of preparation time. Remember that we 
had less than 12 weeks to bring up the new Deep Blue for the 
rematch. We did, out of good will, spend time looking over the 
new chess clock and the document for the new time control. It was 
the team's opinion that there was no way we could program in the 
new time control and finish the match preparation in time. 
However, George Paul from IBM Research was asked to evaluate the 
new chess clock, and see whether there was any way that both of 
Garry's requests could be satisfied. We were relieved when, after 
spending days examining the clock, he reported that it had 
enough problems handling even the regular time control. After 
George's report, Audemars Piguet fixed the clock in time for the 
rematch, with a regular time control. 

Early in the contract negotiation for the rematch, it was agreed 
that, if Garry offered a draw, the operator would be allowed to tell 
Deep Blue of the offer. If the team decided to decline the draw offer, 
Deep Blue would then be told that the draw offer had been 
declined. The purpose of this rule was to prevent the repeat of the 
situation in the 1996 match where Mike Valvo forced the Deep Blue 
team to make a rash decision after Deep Blue had come up with a 
move, without knowing that a draw offer was being deliberated. As 
far as we were concerned, the ability to be able to override Deep 
Blue, and accept draw offers, was a courtesy out of respect for the 
human player. Chess has too many exceptions, and it is not possi- 



The Holy Grail 207 

ble to program the computer to understand that a position is drawn 
in each and every circumstance. Without the ability of the pro- 
grammers to override, the human opponents might be forced to 
needlessly play out drawn endings, which would serve no useful 
purpose other than tiring them. As the rematch drew near, Garry 
had second thoughts, and before agreeing to the new rule, wanted 
to add another rule that the Deep Blue team would not be allowed 
to see how Deep Blue assessed the position while deliberating on 
the draw offer. This was asking for too much as our deliberation 
would then be made without any input from Deep Blue the player. 
The match was supposed to be between Garry and Deep Blue, after 
all. What if Deep Blue finds a brilliant win that no human being 
would see? The match rule for draw offers was never resolved to 
anyone's satisfaction. Internally, the Deep Blue team made the sim- 
ple decision that we would not accept any draw offers in unclear 
positions. In the rematch, he did not offer draws in unclear posi- 
tions, so the problem never came up. 

Since our first ACM game in 1986, ChipTest and its successors 
had played over one-hundred formal tournament or match games 
in public, and Garry had or could have had access to every one of 
them. The match contract explicitly obliged the Deep Blue team to 
provide, when requested, the scores of all our public tournament 
games. These scores were readily available, as we were at the top of 
the computer chess food chain, and our games were always care- 
fully scrutinized. But, in my recollection, Garry never bothered to 
request the game scores. This, however, did not stop him from 
repeatedly complaining that he had no access to any Deep Blue 
games. True, he did not have access to games played by the newest 
version of the program, but usually neither would he have had 
access to games played by, say, the new and improved Anand dur- 
ing the months of training before their world title match. One 
might argue that the newer version of the program was vastly dif- 
ferent from the old one, so Garry should have been given a chance 
to study it. This is a backward argument. The newer version was 
vastly different only because we had put a tremendous amount of 
work into it, just as Anand and Garry himself would have done. 
Should Anand be forced to give Garry sample games just because 
he had improved himself from six months of training? If he had 
asked Anand to give him sample games to study, any reasonable 
person would have considered the request ludicrous. Why should 
he be the only one to have secret weapons for the match? 



208 Chapter 12 

We wanted to be fair, but to give Garry games from our pre- 
match training would have placed us at a serious disadvantage. In 
the exhibition match between him and Deep Thought in 1989, he 
had unleashed a devastating new opening line that busted Deep 
Thought's opening in game two. Deep Thought was dead on 
arrival when it found itself out of its opening book. If Garry had 
some idea of what openings Deep Blue would play, who is to say 
that history would not be repeated? We wanted to play the World 
Champion, not his home preparation against our openings. Could 
we have given him some token game scores nonetheless? 
Theoretically, yes. But there would have been some serious com- 
plications. If he had not been able to take advantage of the game 
scores, then we would have been accused of deliberately mislead- 
ing him. If he had been able to make use of the game scores, then, 
in order to make the match a fair one, we would have had to make 
substantial changes to the opening books, change the program, 
and possibly even build a new machine (so that it could play new 
openings well) 2 . Garry would then ask for new games. Back to 
square one. 

Garry had made similar complaints during the 1996 match, 
but they were not quite of the same intensity. Why did he still 
complain now, when he had already played against Deep Thought 
twice and Deep Blue six times, that he did not have any Deep Blue 
game to study? One possible reason was that, in his mind, he 
deserved to have new games to study, even though it was not a 
right that he would have against a human opponent. There was 
another possibility. During Frederic Friedel's visit to IBM Research, 
he made an interesting remark that might shed some light. Before 
the 1997 rematch, Garry had had great success in top-level chess 
tournaments against other top human players, and appeared to be 
at the peak of his career. One of the secrets of this great success, 
according to Frederic, was that he was revamping his entire open- 
ing repertoire with the help of commercial chess programs. These 
programs were used to point out potential problems in his reper- 
toire, and to find new moves to surprise his opponents. Deep Blue 
was a far more powerful chess machine than any program that 



2 To play certain openings well may require a program to recognize chess posi- 
tional evaluation features that no other chess programs, including Deep Blue, 
recognize. In that case, we would need to add new hardware, or at least new soft- 
ware, so that the machine can play the openings well. 



The Holy Grail 209 

Garry had access to. If he could use commercial chess programs in 
his opening preparations to great effect, a group of Grandmasters 
working together with the help of Deep Blue surely would produce 
some great opening surprises. Was Garry, one of the greatest chess 
opening theorists ever, actually afraid of our home opening prepa- 
rations? 

Garry's problems going into the rematch went beyond just 
general nervousness. For someone nervous about his own chances, 
he, surprisingly, grossly underestimated Deep Blue's strength. In 
the pre-match press conference, a reporter asked him how strong 
he thought Deep Blue was relative to the commercial chess pro- 
grams. He answered that he expected Deep Blue to win eight out 
of ten games against the best commercial chess programs. "Pico" 
Deep Blue Jr, which was about one-thousand times slower than 
Deep Blue, probably would win at least eight out of ten games 
against the best commercial programs, based on the experimental 
data we had. A ten-fold difference in speed between two identical 
programs would very likely produce a winning ratio of 8 to 2, and 
we are talking about a base program that was already capable of 
winning eight out of ten games, speeded up a thousand times ... 
When 1 heard Garry's statement, I said to myself, "Boy, Kasparov is 
in for some shock." 

Before the 1996 match, Garry looked untouchable, and at the 
end of the 1996 match, he was untouchable. Before the 1997 
rematch, however, chinks were already showing up in his armor. I 
liked our chances. 

Setting Up Shop 

The rematch took place on the Manhattan Island of New York 
City, roughly half an hour drive from our lab. In theory, we could 
have commuted from our homes, but most of us stayed in 
Manhattan throughout the entire match. Joel lived in lower 
Manhattan, but stayed with the rest of us in the Michelangelo 
Hotel, which was just across the street from the Equitable Building, 
the match site, in upper Manhattan. Miguel, who arrived from 
Spain just before the match, also stayed in the Michelangelo with 
his wife. Miguel, Nick, and John served as Grandmaster commen- 
tators for the reporters in the pressroom. Miguel was also serving 
as our second when Joel had to give interviews to reporters during 
the games. Nick and John, having fewer responsibilities and being 
native New Yorkers, went home every night. All the Grandmasters 



210 Chapter 12 

were involved in the opening preparation sessions between games. 
IBM set up a customer hosting area in the Michelangelo Hotel, 
with a microwave video feed from the match site. The invited IBM 
customers could watch the games and the live commentaries in a 
comfortable and relaxed setting. Food and wine were served con- 
tinuously. Nick and John, both with a taste for fine wine, were 
frequently sighted at the customer hosting area. After the match, 
Garry made the outrageous claim that they had played for Deep 
Blue from the Michelangelo. Nick and John were just there to 
enjoy the food and the wine, taking care of their gastronomic 
delight, and rubbing shoulders with the IBM customers. 

The decision to host the rematch in New York City created 
some difficult problems early on. The match negotiation started 
more than a year beforehand, but it was already quite difficult to 
find a venue that was of the right size, had adequate facilities, and 
proper vacant time slot. At first, the Millennium Hotel was 
declared as the match site but after the rematch announcement, 
the hotel said that a wedding ceremony was booked for the last 
weekend of the rematch, and wanted IBM to pay an additional fee 
so that the ceremony could be rescheduled. It so happened that 
IBM CEO Lou Gerstner usually hosted his stock analyst meetings 
at the Equitable Building, and quite a number of IBMers were 
familiar with the site. The Equitable Building, the headquarters of 
Equitable Insurance, is a modern building with all the necessary 
wiring for video connections between floors already in place. After 
an extensive examination, the new site was decided to be the bet- 
ter venue, and the Millennium Hotel booking was cancelled. I 
wonder how the managers of the Millennium Hotel felt after see- 
ing the massive publicity, once the rematch got under way? 

The Equitable Building was an imposing skyscraper, as befits 
the headquarters of a big corporation like Equitable Insurance. The 
Broadway Theater district was close by, and during one of the rest 
days, show tickets were secured for the IBM team producing the 
off-Broadway chess show. None of us on the Deep Blue team went 
to the show, though. Joe, Murray, and I had worked for nearly 
three decades, combined, on the project, and we intended to keep 
our eyes on the ball. Broadway shows could wait a while. 

Within walking distance was the Rockefeller Center. The 
Saturday morning before game five, some enterprising IBMers, led 
by Matt Thoennes, pushed the "life-size" Deep Blue model, which 
normally was on the stage at the match auditorium, into the street 



The Holy Grail 211 

right next to the live broadcast of the NBC's 'Today' show in the 
Rockefeller Center. The 'Today' hosts were quite surprised to find 
the big black box outside their window. 

The main entrance for the Equitable Building was on seventh 
Avenue. After passing through the main entrance, the spectators 
arrived in a majestic marble lobby, three stories high, with daylight 
coming in through the glass ceiling. Directly facing them on the 
far wall was a giant painting by Roy Lichtenstein. As a result of 
security concerns following the Oklahoma City and the World 
Trade Center bombings, the audience was asked to pass through 
metal detectors and a security check. They then had to go one 
flight down to the 480-seat auditorium, which served as the com- 
mentary room. Not all the seats were available to the general 
public. IBM, ACM, and Garry all had some reserve tickets. Some of 
the IBM tickets were given to the United States Chess Federation 
for their guests, and most of the others went to lucky school chil- 
dren. A small number of IBM tickets were reserved for invited 
customers who wanted to see the auditorium first hand. Reporters 
were allowed into the auditorium when the fire safety capacity was 
not exceeded. 

At the front of the auditorium, three giant projection screens 
displayed live video feeds of the chessboard, the game room, and 
the computer analysis board used by the commentators. The pro- 
jection equipment was provided gratis by AmPro, the company 
who provided the same service in Philadelphia. There had been 
talks earlier with Matsushita. Had the talks worked out, video walls 
would have been used instead of projection screens, and live 
match video would have been shown on the giant Panasonic/NBC 
TV screen in Time Square. However, some personnel changes took 
place at Matsushita, and the talks broke down. 

Below the projection screens, on the right side of the stage, 
stood a "life-size" model of one frame of an RS/6000 SP supercom- 
puter. The frame was a refrigerator-sized box well over six feet high. 
Garry himself could have fitted nicely into it. A real frame would 
house up to 16 RS/6000 workstations. The frame on the stage was 
actually just a wooden box model of the real machine. After the 
match, this wood model was shipped to California on a charter 
plane for TV commercial shoots. 

Mike Valvo, the match arbiter for the 1996 match, was serv- 
ing as one of the three main commentators in the auditorium. The 
main commentators from the 1996 match, GM Yasser Seirawan 



212 Chapter 12 

and IM Maurice Ashley, were the other two. Of the three com- 
mentators, I thought Yasser was decidedly pro-Kasparov, while 
both Maurice and Mike were relatively neutral. Yasser was the first 
American player to become a World Championship Candidate 
after Robert Byrne (USA) lost to Boris Spassky (USSR) in 1974. If I 
am right about this bias, it was kind of interesting, especially con- 
sidering that Garry refused to share the same stage with him 
during the 1997 match. Yasser and Garry used to be close friends 
in the late 1980s, but chess politics got in the way of friendship. 
Maurice first became well known when Raging Rooks, a Harlem 
junior high school chess team he coached, won the US Junior High 
School Chess Championship in 1991. The first time I met him was 
when his team visited IBM Research shortly after winning the title. 
Besides his coaching activities, one thing that really impressed me 
about him was his stated desire to become the first Grandmaster of 
African descent. Not long after the rematch, his dream became 
reality. Good things do happen to good people. 

The thirty-fifth floor of the Equitable Building, where the games 
took place, was off limit to the spectators. Only one regular eleva- 
tor goes up there, and only people on a restricted list were allowed 
by the security guard to enter it. After taking the elevator, you 
would be greeted by another security guard, who would check your 
name against the list. Since very few people were allowed, after a 
few days the security guard would recognize your face and greet 
you with a smile. 

To get to the game room, guests and reporters alike passed 
through a storage room that was connected to the game room by 
a winding temporary corridor. The corridor was constructed with 
sound proofing material to cut off the noise from the storage room 
air conditioner. The game room had many high-power lamps and 
video equipment that generated lots of heat. The storage room air 
conditioner provided much-needed air conditioning for the game 
room and had to run continuously. On entering the game room, 
you would see a single row of chairs. One of the chairs- -ornate yet 
comfortable — looked quite inviting. This was Clara's chair that she 
had personally chosen. Clara was always present during the games, 
and it was simple courtesy for IBM to make sure that she was as 
comfortable as possible. Beyond the chairs, you could see a parti- 
tioned area that occupied roughly a quarter of the game room. This 
was Garry's rest area, furnished with a sofa, table, fruits and bottled 



The Holy Grail 213 

water. He liked bananas, and every day, fresh, unblemished 
bananas would be delivered to the table. Another soundproof 
room was provided for him on the same floor, stuffed with more 
goodies. There was no port-a-john in his room this time, possibly 
because of the difficulty of bringing one in. He shared the same 
rest room as the rest of the crew on the thirty-fifth floor, except 
that he had the highest priority. Nobody, including members of 
the Deep Blue team, was allowed to use the rest room while he was 
in it. 

Beyond Garry's resting area in the game room was another 
door, usually closed. It was actually the first of two soundproofed 
doors leading to the sound control room. It was also Garry's pri- 
vate entrance. Beyond the two doors was a draped corridor. 
Whenever he opened the first of the double doors, the video crew 
and the IBM team were informed to stay away. The draped corridor 
led to both Garry's private room and the common rest room. The 
only way from the Deep Blue operational room to the common 
rest room was also through the draped corridor. As he again had 
highest priority, this meant that no one on the IBM team was 
allowed to go to the rest room whenever he was out of the game 
room, even when he was not using it. 

The front half of the game room was the playing area. 
Compared to the sterile environment of the 1996 match, it looked 
almost homely. The room was really a TV studio used by senior 
executives of Equitable Insurance for broadcasting and recording 
corporate messages, and was set up to look like the study of an aris- 
tocrat or perhaps a rich lawyer. In the center of the front wall was 
a big poster for the rematch. If there was a painting behind the 
poster, it probably had a medieval scene, perhaps of a knight res- 
cuing a damsel in distress, or maybe a portrait of the patriarch of 
the family in full battle regalia. Several paintings of medieval 
scenes adorned the walls. The bookshelves were full of old law 
books. On the shelves was a ship model as well as several wooden 
duck decoys. Potted plants livened up the well-lit room. 

The lighting was a little too bright for Garry but he compro- 
mised since he wanted to look good on the video. The lighting 
above the playing table had to be adjusted several times until he 
was finally satisfied that there was no more shadow around his 
queen. The playing table was custom made, with a forty-five- 
degree bend at the rear toward the Deep Blue operator on the right. 
A black, $5000 IBM flat panel display sat on this part of the table. 



214 Chapter 12 

To the left, Garry sat on a studded leather armchair of the exact 
firmness he liked. On the right, the Deep Blue operator sat on a 
similar chair. The match arbiter, Carol Jarecki, sat at a desk to the 
right. Murray, Joe and I rotated as the operator from game to game, 
with occasional substitution during long games. The match rules 
forbade the operator from leaving the chair while Garry was think- 
ing, so the operator was usually fairly uncomfortable after staying 
at the table a long while. 

Deep Blue itself was also on the thirty-fifth floor. A few 
months before the rematch, IBM RS/6000 division announced the 
availability of a new RS/6000 SP, which had CPUs (Central 
Processing Units) that were about twice as fast as the CPUs in the 
model we used for the 1996 match. The demand for the new one 
was outstripping the supply. The only way that we could have 
access was to get one that was being tested on the factory floor 
prior to being shipped to the customer. Instead of the usual, the 
RS/6000 SP in question would be doing the floor testing on the 
thirty-fifth floor of the Equitable Building, in front of the eyes of 
the world. 

The RS/6000 SP arrived at the Equitable Building a week 
before the rematch, with all the chess cards in place. It had two 
frames, each frame housing 15 RS/6000 nodes for a total of 30 
nodes, or 30 RS/6000 workstations. Each SP node contained two 
chess cards, with eight chips per card. The total number of chess 
chips in the system was 480. Theoretical maximum search speed 
was about one billion positions per second. This is the speed guar- 
anteed-not-to-be-exceeded by the system. The actual maximum 
speed was around 200 million positions per second. Before the 
move, the people handling the logistics told us that there would 
not be any external phone access to the Deep Blue RS/6000 SP. So 
we all moved into the Michelangelo Hotel the day after the 
machine was moved, since there was no way for us to access the 
machine from our offices in Yorktown, or our homes. 

When we went up to the thirty-fifth floor of the Equitable 
Building, we found Deep Blue ensconced in a small closet-sized 
space at a far corner of the building. There was a dedicated air con- 
ditioner for the machine, but the space around it was quite warm. 
The adjacent space was a little bit cooler, and a few days later we 
installed a floor fan to circulate this air to help reduce the temper- 
ature. The whole machine area was too warm even for someone 
wearing a short-sleeve T-shirt and shorts. We did not have access to 



The Holy Grail 215 

other parts o f the thirty-fifth floor a t the time, and i t was quite dif- 
ficult to work for a long period on Deep Blue, as we were forced to 
stay in the machine area. For a couple of days, I ran extensive tests 
of the system when Joe was not running his own test on the search 
efficiency of the parallel program. Some of the chess cards 
appeared to have been damaged in transport, or perhaps suffered 
from the effect of the high heat, and had to be replaced. The 
machine area and the game room were separated by three other 
rooms, and had different air conditioning systems. It was therefore 
quite funny to hear Garry complaining in public, a year after the 
rematch, that the air conditioner was adjusted for Deep Blue's 
comfort instead of his own. During the rematch, Kasparov did 
complain that his resting area in the game room was too cold. It 
turned out that there was an air vent right above his resting area 
and a deflecting surface for the airflow was installed afterwards. 

Besides the final testing, there was still some work ongoing in 
both the evaluation function and the opening book for Deep Blue. 
A hotel suite was supposedly reserved at the Michelangelo so that 
we could work on the opening book with the Grandmasters. IBM 
had made block reservations at the hotel, and we were quite sur- 
prised to find that our suite had already been taken. For several 
days, we had nowhere to do the opening preparation except on a 
coffee table in a small room next to Deep Blue's room. Finally, we 
figured out what happened to our suite. When IBM was making 
the hotel reservations, Garry wanted to stay in a hotel closer to 
Central Park, so his entourage was all booked into the Plaza Hotel. 
A suite was also reserved for the Kasparov camp at the 
Michelangelo for the duration of the rematch, but not before. 
Garry's new manager, Owen Williams, arrived at the Michelangelo 
a week early, and accidentally took the suite that we had reserved. 
We got our suite when the one reserved for the Kasparov camp 
became available. 

Three days before the rematch, we were finally allowed to 
move into the operations room, which was to become our home 
base. The operations room was probably normally used as an 
editing room as there were panels for switching sound tracks and 
video channels on one side. The room was cramped and during 
the games normally no more than six people could crowd in. Ken 
Thompson, of UNIX and Belle fame, was always present. He was 
serving as the impartial observer for both the camps. During the 
games, Ken was the only person outside IBM who got to see Deep 



216 Chapter 12 

Blue's running log, besides the Grandmasters and the occasional 
reporter. The running log detailed Deep Blue's "thinking" 
process, namely, what it planned to do in various circumstances. 
At least one of the three Deep Blue team members was always 
present, ready to swap positions with the operator at the table. 
Either Joel or Miguel was always around, depending on which 
one of them was in the pressroom. A couple of other IBMers went 
in and out from time to time. Every once in a while, a reporter 
from a major media outlet would be escorted in for an impromp- 
tu interview. A chair in the operations room was probably the 
best seat in the house during the games. There were video feeds 
from the game room and from the auditorium, so you would not 
miss anything important happening in the building. (It would 
have been nice to have had a video feed from the pressroom, but 
none was available.) You had at least one Grandmaster on hand 
for chess discussion, and at least two experts (Ken and one or two 
of the three Deep Blue team members) for computer chess dis- 
cussion. More importantly, you got to see what the best chess 




The Deep Blue group in the game room 
before the start of the 1997 match. 

Standing from left to right: Jerry Brody, Feng-hsiung Hsu, 
joe Hoane, Joel Benjamin, and C / Tan. Sitting: Murray Campbell. 



The Holy Grail 217 

playing entity in the world was "thinking" while all the discus- 
sion was going on. 

We were ready for the show. 

It Saw the Mates! 

Two days before game one, a press conference was held on the 
forty-ninth floor of the Equitable Building. It was originally esti- 
mated that about one-hundred journalists would show up but in 
the end over two-hundred arrived. The room was completely 
packed. ABC, CBS, NBC, CNN, Fox, BBC and a few other European 
TV networks sent their crews. After the press conference, I had a 
chat with an IBM old-timer who had been involved in the product 
launch for the IBM 360 computer in 1964, arguably the most 
important launch in IBM corporate history. He exclaimed, "This is 
far bigger than the IBM 360 launch. There is nothing in IBM his- 
tory comparable to this." As the match progressed, more and more 
reporters showed up to cover it. Eventually, there were simply too 
many reporters, and the pressroom was moved to a much bigger 
room on the fiftieth floor, with a spectacular view of the city. 

In both Deep Blue matches, photographers were allowed to 
take pictures at the start of each game. Now, there were too many 
photographers, and they had to be divided into two batches for 
each game. First one group was allowed to cram into the room, and 
take a few pictures. Then they were ushered out and the second 
group allowed in. Some photographers were caught trying to stay 
on for the second session; there were even some cases of violence 
as the photographers jockeyed to get the best shots. At the start of 
game one, a security guard hired by IBM was punched by a pho- 
tographer who subsequently had his press pass revoked. 

Garry and I smiled broadly and shook hands before we sat 
down for the first game. The rematch had a total prize fund of $1.1 
million, with $700,000 for the winner, and $400,000 for the run- 
ner-up. At the moment that we shook hands, he was probably 
smiling about his $700,000. I was wondering what would happen 
when he received his $400,000 check. We would not get to share 
the $700,000 winner's check — Deep Blue's share would go back to 
IBM. But seeing our dream finally coming to fruition would be 
more than enough. We would all be on cloud nine. All the per- 
sonal sacrifices would be justified. Would some of us shed tears of 
joy? Would I? 

Despite Frederic's claim during his visit that Garry would win 



218 Chapter 12 

easily if he played anti-computer chess, but would never do it — he 
did just that. For the entire first game, his pieces never left his half 
of the board. 

After the first nine moves, Garry and Deep Blue arrived at a 
well-known position that had been played many times at the 
Grandmaster level. The position was closed and highly positional, 
exactly the type of position that was supposedly hard for a com- 
puter. Garry next played an unusual move, 10. e3, moving his 
e-file pawn forward one square. A more normal move was to move 
the e-file pawn forward two squares, offering a pawn trade. He 
wanted to keep the position closed for a little bit longer. With the 
move e3, he revealed his strategy for the entire match — play anti- 
computer chess. We were about to see one of the strangest chess 
matches played at the highest level. Why did he decide to adopt 
this strategy? How and for how long did he prepare for this style 
of play? I have a theory. 

Frederic believed that Garry would win easily if he played 
anti-computer chess. Frederic and Garry were close friends, so he 
must have known about this opinion. We had told Frederic during 
his visit that Deep Blue had a good chance of winning the match 
and it is reasonable to assume that he relayed this information to 
his friend. I don't know whether Garry had begun his match 
preparations before receiving this information; he was not particu- 
larly forward about things related to his training. Before the start 
of the rematch, he had stated publicly that he had prepared seri- 
ously for it. Afterwards, he stated that he had not prepared for it. I 
believe that he did prepare for the rematch, at least in the sense of 
preparing to play anti-computer chess. Furthermore, I believe that 
Frederic's visit to Yorktown might have triggered Garry's decision 
to play in this way. One possible scenario was that the prematch 
confidence of the Deep Blue team might have caused him to won- 
der whether we had prepared some killer opening lines. Well, our 
team of Grandmasters, together with Deep Blue, found quite a few 
opening innovations, but they were not necessarily killer lines 3 . 
However, Garry did have reason to be cautious. The easiest way for 

1 Deep Blue found some of the new lines, including one that was rejected by the 
Grandmasters, in the Queen's Gambit Accepted opening, it found a pawn sacri- 
fice (e4) for White that it liked. The Grandmasters thought that Deep Blue's 
move was too ambitious as the compensations for the pawn seemed to be tenu- 
ous. A few months after the rematch, the same move was played at the 
Grandmaster level, and with good results. 



The Holy Grail 219 



him to avoid our new opening lines was to play anti-computer 
chess by abandoning normal openings altogether. 

Assuming that Garry decided to adopt the anti-computer style 
of play after Frederic's visit to Yorktown, then he had just over two 
months to prepare the new style. Since he did not have access to 
Deep Blue, the commercial chess programs were most likely his 
sparring partners. Given his statement at the press conference that 
Deep Blue should beat the commercial programs by 8 to 2, my 
guess is that his success rate against the commercial program must 
have been better than 8 to 2. Of course, he was quite wrong in his 
assessment of the relative playing strength of Deep Blue and the 
commercial programs. 

After Garry's e3, Deep Blue seemed to be in a like mind, and 
moved its h-file pawn one square. It was a slight error but, consid- 
ering his last move, it was playable. A better move would have 
been to develop the queen. This move and Deep Blue's next move, 
which placed its queen in an awkward position, pointed out a 
problem that Deep Blue had. When the new chess chip was being 
designed, little design effort was put in the chess hardware on how 
to position the queen. The chess books that I had, only gave cur- 
sory treatments on the positional play of the queen. After this 
game, Joe wrote some simple code on where the queen should be 
placed. 



ii m i! 

ill Hi. 



abcdefgh 

Position after move 14. 



220 Chapter 12 

For the next few moves, Deep Blue played some other ques- 
tionable moves, but given the fact that Kasparov was playing 
suboptimal moves himself, Deep Blue's position was still tenable. 

After move fourteen, Deep Blue advanced both the g-file 
pawn, which was in front of its own king, and the e-file pawn to 
the fifth rank. Deep Blue's king was almost completely devoid of 
pawn protection. The f5 square now became a weak square — it 
could no longer be guarded by the adjacent e-pawn and g-pawn. 
The position looked horrible even to my eyes. The question was 
whether Garry could exploit the f5 square easily. When I looked 
more closely at the position, it dawned on me that maybe Deep 
Blue's treatment of the position was fine, even though it was cer- 
tainly not one that would be chosen by a human player. For the 
next ten moves or so, Garry tried very hard to get a knight to the 
f5 square, to no avail. In the 1996 match, even though Deep Blue 
won the first game, I had the distinct feeling that Garry was the 
one calling the shots. As I watched this game unfold, I had a com- 
pletely different feeling. It felt more like a negotiation. He was 
trying to execute his plan, and Deep Blue kept on finding moves 
to thwart it. He then found a new way to continue his plan, and 
Deep Blue came back with another counter plan. Neither side was 
getting what it wanted. Deep Blue, however, would never get frus- 
trated. I had no idea how the game would end, but I did know that 
if every game where Garry had White was going to be like this one, 
he would be in deep trouble. 

Deep Blue's move 22. ... g4, which resulted in its g-file pawn 
being traded with Garry's h-file pawn, was controversial. Some 
analysts considered it justified in that Deep Blue gained better 
piece activity; since Garry did not get a clearly winning position as 
a result, it was hard to say for sure that it was a bad move. The 
move was the result of a bug we had introduced just before the 
rematch. An automatic tuning run for the evaluation function 
pointed out to us that the weighting for one class of king safety 
terms should be increased. When we increased the weights of the 
terms, Deep Blue appeared to play better, and we left the new 
weights in. What we did not realize was that, in extreme cases, the 
new weights reached the maximum allowed value, and became 
saturated. In other words, Deep Blue no longer distinguished 
between a very bad position, and an even worse position. Tossing 
away the g-file pawn therefore meant nothing to Deep Blue, as 
long as it eliminated Garry's h pawn. We found the bug only after 



The Holy Grail 221 



8 

7 
6 
5 
4 
3 
2 
1 



i ri*y a j 



W 



"^^wm 



:; # Ml IP id 

a m a iix 

hah 1a "ha 



^pX*"^ & 









St-Ji.i.*./* /s *,*,/, -^//.y././ , fj>,sst/f 



abcdefgh 

Position after Deep Blue's move 22. ... g4. 

game four, when a similar problem showed up. In both cases, it 
was far from certain that the moves were bad. 

Deep Blue's move 28. ... f5 put the game into a crucial phase. 
At the board, both Garry and I were taking it calmly. He was prob- 
ably expecting it, even though the move was played at the exact 



8 
7 
6 
5 
4 
3 
2 
1 



"Up psT^i : 



"AW 



9ttv, ^ :$%%$ ■mm, £™^ 

- ■"■ Si Hi. 



ifi II A tti m 



V^vmiF* 1 **/, 



mM 



y /A 






O » IP II 

^=#/ ^p?^ %^% ^%% 



abcdefgh 

Position after Deep Blue's move 28. ... f5. 



222 Chapter 12 



instant of his twenty-eighth move. In the auditorium, the com- 
mentators were shocked — none of them expected it. It was 
probably the best move in the position, and Garry had the choice 
of whether to make an exchange sacrifice (giving up a rook for a 
bishop or a knight) or not. Deep Blue was expecting him to make 
the exchange sacrifice, although it assessed the position after the 
sacrifice to be much better for Garry. Post mortem analysis indi- 
cated that he would have been much worse off if he had not made 
the sacrifice. 

As it was, Deep Blue did not like its position. Usually, this 
meant that I would be feeling miserable, but I wasn't — the position 
was messy and yet we would have our chances. In the auditorium 
and at the pressroom, people were starting to talk about an upset 
win for Deep Blue, but they were not seeing Deep Blue's output. 
Garry quickly made the move that initiated the exchange sacrifice. 
The next few moves all went as Deep Blue expected. Both Garry 
and Deep Blue apparently had seen all the ramifications. 



8 
7 
6 
5 
4 
3 
2 
1 




abcdefgh 

Position before Deep Blue's move 33. ... Qb5. 

When the dust settled, however, Deep Blue played a ques- 
tionable move, 33. ... Qb5, which allowed Garry to play Qfl to 
trade off the queens. Deep Blue was willing to trade the queens 
because its king was less safe. But Garry also had two connected 
passed pawns (the f5 pawn and the g6 pawn) that could become 
very valuable in the endgame phase if the queens left the board. The 



The Holy Grail 223 



ability to express this endgame feature existed in the chess hard- 
ware but was not in use. Joe wrote the code to use this hardware 
ability after game one, but the new code was not active until we 
had a rest day after game two to check it out. Afterwards, Deep 
Blue played 33. ... h5 instead. 






Mil 

m W%. <-j • 



life 

M fH H: o ^ * 

48 MA"" ' 



m-jm 
4 



& 



<■><■. 



•i" 



% 






abcdefgh 

Position before 36. ... Kf8. 36. ... Ng4 was the last chance to draw. 

The move 33. ... Qb5 was probably bad, but it was not neces- 
sarily losing, as Deep Blue had a drawing move later on move 
thirty-six. Instead of 36. ... Kf8?, 36. ... Ng4! appears to draw. Post 
mortem analysis with Deep Blue suggested that Deep Blue might 
have played 36. ... Ng4 if it could have searched one more ply, 
which would have required roughly a six-fold increase in search 
speed. 

Without 36. ... Ng4, Deep Blue's position quickly went down- 
hill. By move forty-four, we were close to resigning for it. Then it 
suffered the same bug that caused Deep Blue Jr to toss a pawn 
against Christiansen in its training match. Instead of playing 44. ... 
Rf5 and prolonging the game as it intended, it played the instant- 
ly losing move 44. ... Rdl, and after Garry's reply, we immediately 
resigned. That night, Joe worked late and finally fixed the bug that 
had haunted us for over a month. 

After the game, we knew we had work to do, but the team was 
in high spirits. We had Black, and yet Garry's anti-computer chess 
experiment almost backfired on him. Deep Blue mishandled the 



224 Chapter 12 

position once the position became double edged but, if this sce- 
nario continued, we would have our share of wins as Black. For a 
group of people who had just lost the first game of a six-game 
match, we must have looked surprisingly cheerful to the reporters. 
Garry, meanwhile, apparently was puzzled by Deep Blue's last 
move in the game. The day after game two was a rest day, and 
when I surfed the web to catch up on the match news, I came 
across an article written by Frederic Friedel about what happened 
in the Kasparov camp after game one. Garry was perplexed by the 
move 44. ... Rdl. Deep Blue played it as a result of a bug but Garry 
did not know that. So the whole Kasparov camp went into a very 
deep analysis on why the alternative move 44. ... Rf5 was no good. 
In the end, they concluded that the reason why Deep Blue did not 
play 44. ... Rf5 was "It probably saw mates in twenty or more 
[moves]." I could not help but burst out laughing 4 . 

A Bad Move That Wins 

The morning before game two, the team debated whether we 
should make changes to Deep Blue's evaluation function. Joe had 
already fixed the bug responsible for the last move in game one, 
but there were two other changes that we needed to look at. The 
proposed change to the passed pawn evaluation was too drastic, 
and we decided to leave it out for game two. Joe's change to the 
queen evaluation was relatively minor, and we decided to keep it 
in the program. The program that played game two was essential- 
ly the same program that played game one. 

Murray was the starting operator for the day, so I spent time 
in the operations room during the early part of the game. Deep 
Blue opened with 1 . e4. Garry, instead of replying with his normal 
Sicilian 1. ... c5, played 1. ... e5, just as he did in game five of the 
1996 match. Had Garry played 1. ... c5, we would have been ready 
to battle him with the main lines of the Sicilian instead of the c3 
variation that we played in 1996. 

The game quickly became a Ruy Lopez opening (also known 
as the Spanish Torture), named after a 15th-century Spanish priest. 
Nick and Joel, rather than Miguel our Spanish contingent, had pre- 
pared the basic lines for the Ruy Lopez. As the opening quickly 
unfolded, Joel was in an unusually upbeat mood. He had played 

4 One report stated that the Kasparov team also burst out laughing when they fin- 
ished their analysis on why 44. ... RfS was no good. 



The Holy Grail 225 



many games in the Ruy Lopez against Deep Blue Jr and was quite 
confident of Deep Blue's ability to handle the position. 

Meanwhile, in the auditorium, the commentators were saying 
how terrible Deep Blue's opening preparation was. Deep Blue had 
allowed Garry to achieve a closed position. The commentators had 
written off Deep Blue even before the opening phase of the game 
was completed. It was not said, but it was clear that they were 
expecting Deep Blue to lose the game. 



8 
7 
6 
5 
4 
3 
2 
1 



If 



"SSvS?" 






Ji 



v&m. 



mw m 



W/, ~*mW& '/,*»; 

mm. m 



m 




abcdefgh 

Position after Deep Blue's move 16. dS. 

Garry was probably quite happy when Deep Blue played 16. 
d5, blocking up the center, ostensibly giving him the closed posi- 
tion that he so desired. Joel was also quite happy. Garry's opening 
selection had given Deep Blue a significant advantage in space that 
could haunt Garry for the rest of the game. Joel had been on the 
receiving end of many punishing games with the same theme. In 
the operations room, slowly but surely we saw Deep Blue's assess- 
ment of its positional edge creeping up. The commentators in the 
auditorium continued to say that Deep Blue was in trouble. 
Grandmaster Yasser Seirawan: "This isn't the kind of position that 
the computer can play very well." International Master Mike 
Valvo: "From the computer's point of view, there's nothing really 
clear in this kind of position." 

Joel, possibly sensing a chance to needle Yasser, whom he had 
known for years, asked to be excused from the operations room so 



226 Chapter 12 



that he could get to the auditorium stage to defend Deep Blue's 
opening preparation. Almost immediately after he left, the com- 
mentators, having had more time to digest the board position and 
seeing Garry shaking his head a little bit, had a change of heart. 
Yasser was starting to say that he liked White. 









abcdefgh 

Position after Deep Blue's move 24. Red. 

Deep Blue's "human-like" move 24. Reel forced its opponent 
to make a hard decision. Garry's 24. ... c4 closing up the c file was 
criticized after the game as removing his own last chance of count- 
er play. Deep Blue was ready to play c4 itself and bust the whole 
position wide open and Garry's move at least stopped that. Garry 
probably did not want to allow the complications and, given the 
fact that Deep Blue had the more active pieces, it was an under- 
standable decision. 

Joel reached the auditorium stage at this point, a little too late 
to needle Yasser. Joel's first remarks were, "Well, I was following 
your commentary from upstairs, and you were talking about how 
closed the position was for so long. I was wondering when you 
guys were going to get around to the fact that the position was bet- 
ter for White, and you eventually did." The audience laughed. 

In the operations room, Joe and I exchanged a look when 
Deep Blue doubled up its rooks on the a-file, and yet refused to 
open up the a-file immediately. Both of us were thinking about 
what happened many months ago after Joe knocked on my door. 



The Holy Grail 227 



The new evaluation feature about rooks on potential open files was 
in play here. I smiled and said to Joe, "Not bad for an afternoon's 
work." The two White rooks on the potential open a-file tied down 
Garry's heavy pieces to his back rank, and forced him to be com- 
pletely passive. 



8 

7 
6 
5 
4 
3 
2 
1 




y //A 



4UJHI 



l«Ai! m 

m m m m 

"■a www/km 

t£aM llH ^ %Mk 



abcdefgh 

Position after Deep Blue's move 26. f4. 

Deep Blue's 26. f4, opening up a second front on the king 
side, was probably partially a result of the new evaluation feature. 
Anatoly Karpov was watching the game live in one of the Internet 
chess clubs. He was making quite a few comments early in the 
game, and was fairly dismissive of Deep Blue's play. Deep Blue's 26. 
f4 did not escape his criticism; although, to the best of my knowl- 
edge, none of the other Grandmasters shared his view — in fact, 
they universally praised the move. As the game progressed, 
Anatoly made fewer and fewer comments, and towards the end fell 
silent. 

In the game room, Murray signaled that he needed a replace- 
ment. Garry was still at the board thinking, so I took the long 
tortuous route through the normal entrance into the game room. 
I sat down with the other spectators and waited for Garry's move 
before making the switch with Murray. When I was in the opera- 
tions room, Garry looked unhappy but otherwise normal on the 
video monitor. 1 was taken aback when 1 looked at him in the game 
room. He looked like someone who had just woken up from a bad 



228 Chapter 12 

nightmare. Part of his face was visibly red as if he had slept on that 
side for a while. 

Soon after I took over as the operator, Deep Blue played the 
two moves that would cause Garry a lot of grief for the rest of the 
match. They were the moves 36. axb5 and 37. Be4. Right before 
axb5, it played the move Bxd6, which was followed by Garry's own 
Bxd6. While computing for the move Bxd6, Deep Blue's evaluation 
dropped. As a result, it went into the panic mode and ended up 
spending just slightly less than 15 minutes on its move thirty-five. 
As you will see later, Bxd6, which was an obvious move, used up 
more computation time than the next two moves combined, 
despite Garry's later claim that Deep Blue spent more time on 36. 
axb5 than any other move played in the match. (He might have 
meant 37. Be4, but Deep Blue spent only a normal amount of time 
on it.) 

While writing the previous paragraph, 1 saw the log 5 for the 
move Bxd6 for the first time since the match itself, and it showed 
something quite interesting. 





m ii i 



abcdefgh 

Position before Deep Blue's move 36. axb5. 



s At the time of this writing, you can find the game logs by following the links to 
the Deep Blue match site and then to the individual games from the web site 
www.chess.ibm.com. It is possible that IBM may not keep the site up 
indefinitely. 



The Holy Grail 229 

First, when Deep Blue finished the panic mode search for 
Bxd6, it was planning to play 36. Qb6 as the follow-up to Bxd6, but 
only barely, because it already saw hints that Garry might be able 
to do something nasty. 

Before Deep Blue went into the panic mode for Bxd6, it had 
come up with an unusual principal variation when it finished its 
9-ply search. The main line it gave was 15-ply deep. This number, 
however, was only the search depth on the master workstation, 
and did not include the additional search depths on the slave 
workstations (2+ plies), and on the chess chips (4+ plies). The main 
line was really at least 21-ply deep and maybe well over 30-ply 
deep, since there were additional search extensions on the work- 
stations and on the chips. Deep Blue was expecting Garry to give 
up two pawns and push his e-file pawn for a potential attack. The 
score was, however, only slightly more than a pawn up for Deep 
Blue, signifying that it thought Garry would have slightly less than 
a pawn worth of positional compensation. 

At the end of its 10-ply search, it was expecting Garry to sac- 
rifice three pawns, but its score had dropped further to only a pawn 
up. At the eleventh ply, the score for Bxd6 dropped again, which 
then triggered Deep Blue to go into the panic mode. The principal 
variation once again had the three-pawn sacrifice by Garry, but it 
now showed him recapturing a pawn in the process. The line it 
gave was now 18-ply deep, or was really at least 24-ply deep and 
maybe even over 40-ply deep along the deepest line. I remembered 
that I was saying to myself at this point during the game, "Perhaps 
Be4 blocking the e-file pawn would have been a good move some- 
where around here." 

Meanwhile, unbeknownst to me, Yasser gave the two-pawn 
sacrifice line to the audience at the auditorium and over the 
Internet. Deep Blue had discovered the same sacrifice line during its 
9-ply search but had since rejected it as an inferior line for Black. 

After Garry made his move thirty-five, all the other comput- 
ers in the world and the audience in the auditorium were 
expecting Deep Blue to play 36. Qb6, winning a pawn in the 
process. But now Deep Blue even more disliked the compensation 
that Black could get against Qb6. This time it went into the panic 
mode in a different way. When its normal search time was up at 
the start of the 11 -ply search, the expected principal variation was 
falling apart somewhere deep down in the search. It entered a 
temporary panic mode that forced it to resolve the score for move 



230 Chapter 12 

Qb6. When this search was completed, the score had dropped to 
48 points (slight less than half a pawn) 6 . Since this score was more 
than 15 points lower than the expected score 74, Deep Blue con- 
tinued to stay in panic mode to examine the alternative moves. 
The next move completed was axbS, which had a score of 63. This 
score was good enough, so Deep Blue immediately played 36. axbS, 
spending about 6.5 minutes on the move. It was still planning to 
play Qb6 on the next move. Garry replied with axb5. 

On the thirty-seventh move, the score for Qb6 at the eleventh 
ply dropped to 32 points, 84 seconds into the search. After 126 sec- 
onds into the search, Be4 was scored as 37 points and replaced Qb6 
as the best move. The 11 -ply search was completed after 182 sec- 
onds. The search was terminated at its normal time after 199 
seconds. Deep Blue did not go into panic mode on its 37th move, 
it just spent its normal allocation of time. As mentioned earlier, 
this last fact would become interesting, as later events showed. 

It is hard to say with any certainty what causes a chess 
machine that searches two-hundred million positions per second 
to play a particular move. For the moves in question, 1 believe that 
two things came into play. Deep Blue was searching very deep 
along the critical lines, and its king safety evaluation had much 
greater weight than in other programs. 

For 36. axb5, the decision was made because of very deep 
calculations. It turns out that 36. Qb6 might be objectively the 
better move, as White reserves the option of getting back into 
positions similar to the game continuation, while giving Black a 
chance to lose immediately. Based on the post game comments 
given by Garry, it was possible that Qb6 would have ended the 
game almost instantly since he did not see the strongest line. 
But then we would not have the very interesting finish of game 
two. 



6 For the chess enthusiasts, the three-pawn sacrifice line starts as follows: 36. 
Qb6!? Qe7! (the alternative two-pawn sacrifice move Rd8 is refuted by 37. Be4!) 
37. abS Rab8. And now the best line for White is 38. Qe3! abS 39. Be4, spurning 
the three-pawn sacrifice to reach a position similar to the game. If White decides 
to accept the three-pawn sacrifice, then we have 38. Qa6 e4! 39. Be4 (39. Rel Qe5 
40. Re4 Qh2 41. Kf2 Bg3 42. Kf3 Bd6 draw) Qe5 40. Bf3 Rd8 and Black would get 
significant compensation. For example, if 41. Qa3, Re8 42. Re2 Qh2 43. Kfl Re2 
44. Be2 Re8 45. BhS Qf4 46. Bf3 Re3 47. Qb2 Qh2 draws. If 41. Ra3, Qe3 42. Kfl 
Qd3 43. Be2 QfS 44. Kgl QeS 45. Kfl (45. Bf3 Re8) - Re8 46. Bf3 Bc7! 47. d6 Bb6 
48. d7 Red8 49. Rel Qd6, and Black might be winning. 



The Holy Grail 231 

There was some evidence that the king safety evaluation was 
involved in the move decision for move 37 Be4, which was more a 
positional than a tactical decision. After game four, we found and 
fixed the king safety bug that occurred earlier in game one. Deep 
Blue stopped playing Be4 after Joe made the first bug fix, which 
effectively toned down the king safety term. When it was turned 
up again, Deep Blue again played Be4. Since everyone seemed to 
think that Be4 was a great move, Joe decided to use the second bug 
fix, which had the king safety evaluation roughly back at the same 
level as in game two, for games five and six. 

I was not really too happy about the situation when Be4 was 
played. Deep Blue assessed the position as only slightly advanta- 
geous. It was a good thing that Garry did not seem able to read my 
face. My mood improved somewhat when, over Garry's next few 
moves, Deep Blue's evaluation started to creep upward. I was 
expecting the game to last quite a while. Then, suddenly, after 
Deep Blue's forty-fifth move, Garry offered his hand and said, "I 
resign." 1 was surprised as it seemed a little early to resign, but he 
apparently had had enough Spanish torture for one day. Deep Blue 
scored the position as about 160 points up. It was a winning edge, 
but the game was not really over. Garry probably started the game 
in an optimistic mood, especially when Deep Blue closed up the 
center. From then on, he had been forced to play defensive moves. 
The pressure just kept on building, and finally he could not take it 
any longer. 

Garry did not say anything else after he resigned; there was no 
post-mortem discussion this time. He left the building quickly 
with his entourage. We went on to the auditorium stage to a round 
of applause. We were exuberant. Our labor for the last year had 
finally been recognized by the world. Murray, Joe and I had argued 
from time to time in the past, but for the last year or so we had put 
aside our differences and worked on the common goal. We had 
come together with the Grandmasters to be a real team, and all the 
teamwork had finally paid off. Personally, it was a very satisfying 
moment. The six months of work that I had spent on the new 
chess chip was completely justified. For those six months, I had 
poured my body and soul into the project. Having no weekends for 
six whole months, driving home late at night and waking up sev- 
eral times a night to check the simulation had all suddenly become 
worthwhile. We were happy. We felt like proud parents who had 
just seen their new baby taking the first significant step. 



232 Chapter 12 

Yasser was not giving Deep Blue much credit at the beginning 
of Game two but by the end he changed his mind: "I would be 
proud to have this one." Maurice Ashley: "It's a gorgeous game." 
Joel Benjamin had the following to say: "This is a game that any 
human grandmaster would be proud to have played for White. 
This was not a computer-type game. This was real chess." It was the 
best game ever played by a computer. 

Yet, there was a fly in the ointment. 

1 woke up early the next day, which happened to be a rest day. 
Still excited from the last game, I went into the operations room to 
catch up on the Internet news. One of the sites that I visited was 
Mark Crowther's "The Week In Chess". I had had some communi- 
cation with Mark a few years back. He felt that computers would 
destroy chess, and a loss by the World Chess Champion to a com- 
puter would be very bad for the game. I suggested that 
man-machine chess matches could be a great boon to chess, and 
the eventual loss by the World Chess Champion to a computer 
would actually be a great opportunity for the chess world. The 
huge publicity would draw people to the game, and if the chess 
world could respond positively and take advantage of the publici- 
ty to promote the fun and enjoyment of chess to the new 
audience, there could be a massive increase in the game's popular- 
ity. Mark was not completely convinced. Perhaps he agreed with 
my logic but was not sure about how the chess world would react. 

The first statement on Mark's web site about game two sur- 
prised me. He made a blanket statement that game two should be 
a draw. My first reaction was, "Mark is in denial." I took a cursory 
glance at Mark's analysis of the final position in game two, and 
went on to other news items before I went back to the hotel. Over 
breakfast, I mentioned to Murray that Mark was claiming that 
game two should be a draw. While talking with Murray, a thought 
occurred to me, "What if I am the one in denial?" I decided to dou- 
ble-check the analysis given on the web after breakfast. 

I went back to the operations room and tried out the vari- 
ations given on the web with Deep Blue Jr, as Deep Blue was 
being used to test out our opening preparations. Within twenty 
minutes, I was in shock. Yes, it looked like the position might 
be a draw with best play by Garry. There was a way for White to 
avoid an outright repetition draw, but the resulting ending, 
even though better for White, might not be winning. Deep 
Blue's last two moves in the game turned out to be mistakes. 



The Holy Grail 233 



8 

7 
6 
5 
4 
3 
2 
1 




abcdefgh 

The final position in game 2. 

Move 44. Kfl, moving king from gl to fl in response to Garry's 
Qb6+, even though still winning, made life tougher. A better 
move would have been 44. Khl, which wins easily as there was 
no chance of a repetition draw for Black afterwards. Deep Blue 
played 44. Kfl because it thought that it could force the queens 
off the board, and the ending would have been better with the 
White king closer to the center. The real culprit was the last 
move of the game, 45. Ra6, which allowed Black to either get a 
slightly inferior ending or a very deep repetition draw. I later 
discovered that the drawing line was first found by painstaking 
group efforts involving both human chess players and comput- 
er chess programs. Black's best move is 45. ... Qe3, which is 
fairly easy to see from a human perspective, but some of the fol- 
low-up moves for Black are not so easy for human beings. Garry 
obviously did not see the follow-up moves and hence the resig- 
nation. 

After 45. ... Qe3, White could play either 46. Qd7+ or 46. 
Qxd6. To refute either of the moves, you need to do fairly deep and 
tricky analyses. Deep Blue Jr had the critical moves in its main line 
within minutes, but to sort out all the ramifications took close to 
an hour. The deepest line was about 30 plies deep. Deep Blue, in 
theory, should be able to cover 30 plies of forcing lines in real time. 
But there was one huge complication. Deep Blue, like most other 



234 Chapter 12 

chess programs, used the so-called transposition table. The transpo- 
sition table allows a chess program to avoid searching the "same" 
position twice when the new position is a repeat of an old position 
already searched. The problem is that the new position is not real- 
ly the same as the old one. They have different move histories. 
Normally, this fact does not matter, but when repetitions are 
involved, a program could either miss a repetition or find a false 
repetition because of the use of the transposition table. Sometimes, 
when the program searches deeper, the problem eventually goes 
away, and sometimes, the problem never goes away. Anyway, 
either Deep Blue was not searching deep enough, or it was side- 
tracked by the transposition table from seeing the potential 
repetition draw. 

From the pure chess point of view, Deep Blue's last move in 
game 2, Ra6, deserves to be annotated with a "??" mark (this 
means a highly dubious move), as it could have allowed a win to 
be turned into a draw. However, from the psychological point of 
view, the move deserves to be annotated with a "??!!" mark (!! 
means an excellent move. ??!! would mean something like "a high- 
ly dubious but excellent move"). It did win the game, and in light 
of what happened subsequently in the match, some might even 
say that the move won the match. 

Tempest In a Teacup 

I phoned Murray the moment I realized that game two could have 
been a draw. We both felt a letdown. On the other hand, a win is 
a win, and it would be quite interesting to see how Garry would 
respond to the news. I knew that I would not want to be the one 
to tell him. As for the Deep Blue team, we had it easy. Our silicon- 
based player could not care less. 

Frederic Friedel wrote an article on how the Kasparov camp 
found out about the potential drawing lines and how Garry 
responded to the news. Frederic had a flare for the dramatic so I 
always took what he said with a pinch of salt, especially when it 
was related to Fritz, a commercial chess program that he sold. 
Anyway, his account did suggest a few things that were missed by 
other reports. 

According to the article, Frederic worked late after game two 
using Fritz to go down the potential drawing line and analyze the 
resulting positions. He found the drawing moves with the help of 
Fritz, but without knowing what had already transpired over the 



The Holy Grail 235 

Internet. However, at this point he was not too sure whether the 
position was indeed a draw. This was not a big surprise. The lines 
that lead to repetition draws were simply too deep, and Fritz, not 
surprisingly, could not see the repetitions starting from the final 
position of game two 7 . 

The next morning, he asked Yuri Dohokian, Garry's match 
second, to come to his room to look at the analysis. Frederic: "Yuri 
drew a sharp breath at 46. ... Re8 and said, 'We didn't look at this 
move.' He started checking my analysis and after some time came 
to the shocking conclusion: this move really draws." It is not clear 
from the above statement whether Yuri's "we" included Garry him- 
self. But suffice it to say that Garry missed at least one critical 
drawing move during the game and possibly during his own post- 
mortem analysis when he had all the time that he needed. 

Of Garry's reaction, Frederic wrote, "Garry clutched his head 
and froze in the middle of the Fifth Avenue. There were no exple- 
tives, no cursing, just stunned silence." They were on their way to 
an Italian restaurant and Frederic described what happened in the 
restaurant, "... it must have been a miserable meal. It wasn't. And 
he's amazing. After about ten or fifteen minutes, he suddenly was 
fired up and started telling us about some Russian movie." So, 
according to Frederic, Garry took the news well. 

While the news of the potential draw created a crisis in the 
Kasparov camp, among the Deep Blue team members, the main 
topic before game three was whether Garry would continue to play 
anti-computer chess. Our first feeling was that he would start to 
play normal chess because his anti-computer chess tactics were not 
doing too well. He did win game one, but Deep Blue could have 
secured a draw close to the end of the game. From our point of 
view, Garry was a little lucky that Deep Blue did not mix it up fur- 
ther and possibly even win the game. Game two was an outright 
disaster for him. He could have secured a draw as well but, like 
Deep Blue in game one, he did not see the drawing line either. 
Anyhow, the seed of Garry's game two loss was clearly his desire to 
play anti-computer chess, which led him to reach awful positions 
that were supposedly good against other computers, but turned 
out to be dismal against Deep Blue. But this was an analysis based 

7 It is not out of the question that some other program might claim that it sees 
the repetitions, but chances are that if it does, it would be due to false repetition 
claims resulting from the use of the transposition table. 



236 Chapter 12 



on pure chess merit. Our amateur psychological analysis of Garry's 
mental state led to a very different conclusion. He had a very 
strong will; most people would even say that he could be stubborn 
at times. He came to the match with a set plan, and that appar- 
ently was to play anti-computer chess. It was most likely that he 
would not change his plan unless it proved to be a complete fail- 
ure. A good question to ask, now that we are looking back, is what 
would have happened if he had played normal chess for the entire 
rematch? I don't know the answer, but I do believe that it would 
have been a close match as well, and I would not be too surprised 
by a Deep Blue match win. 

By the time game three started, the entire Deep Blue team 
believed that Garry would stick with anti-computer chess no mat- 
ter what. So when he played 1 . d3, it came as no surprise to us. Joe 
was the first operator for this game. The move 1. d3 probably had 
never been played at the highest level of chess. Ironically, it was a 
move that we had used against another chess program back in 
Deep Thought days. The commentators in the auditorium were 
quite shocked by the move. Mike Valvo: "Oh, my God." Maurice 
Ashley: "A cagey move. A shock of shocks in this match. This 
match has everything." Yasser Seirawan: "I think we have a new 
opening move." 

During the first few moves, Garry appeared to be trying to 



8 
7 
6 
5 
4 
3 
2 
1 




™* — - 



HlWiBlWl 
. m m i 



. -'AWL . -^ 



y //////a m 



% 



b m&wm 



/**,,*/* 



W/..^ y //// 

W 



7>rrr!?. A— -f. .ZVttt'Z 






>Wi & P 



abcdefgh 

The position before 4. ... d6. 



The Holy Grail 237 

trick Deep Blue into playing the White side of the Sicilian defense, 
but, with him having an extra move, which could be critical in 
some of the Sicilian lines. Deep Blue's 4. ... d6, instead of the risky 
d5, sidestepped this possibility. 

After Garry's 5. Nc3, the game had transposed back into a 
known English opening position. In this position, British 
Grandmaster Michael Adams once played Be7 as Black. Deep Blue 
was given a hefty bias to play 5. ... Be7 as a result of this one sin- 
gle game played by Michael. Many commentators criticized the 
move as a little bit passive. After this game, Murray changed the 
code used to compute the move bias in known positions. Michael 
was a very good chess player, but giving Be7 a hefty bias just 
because he had played it once, seemed excessive. 



8 
7 
6 

5 
4 
3 
2 
1 




-3 



m 



il Wi.il 

y ff ■ ii 



&m& 



'//***: 



Tfl 








abcdefgh 

The position after move 14. 



After Deep Blue's 14. ... c5, Garry had a slight edge but, 
according to the Grandmasters who had had time to think about 
the position after the game, Deep Blue's position appeared to be 
solid. Some Grandmasters had the first impression that the posi- 
tion clearly favored Garry. The British Grandmaster John Nunn 
commented, "True, Deep Blue has played the opening in a way 
which appears very odd to human eyes, because it doesn't fit any 
of the normal patterns for playing with the English opening. 
Ideally, White would like to set his kingside pawn in motion, but 



238 Chapter 12 



8 

7 
6 
5 
4 
3 
2 
1 



wi'iw' 

m M Hill 




m ■ 



^■W;-?-.*— * 



mmmM 






a saa I 







abcdefgh 

The position after move 22. 

it isn't easy to achieve this." Sometimes, searching two-hundred 
million positions per second produces unusual new solutions. 

The first critical moment of the game came after Deep Blue 
played 22. ... Qa5, and Garry replied 23. Bd2. At this point, Deep 
Blue had two choices, to play Qxa3 or Nxg4. The machine at first 
preferred to play Nxg4, but apparently Garry would get too much 
play, so it played Qxa3. This move wins a pawn, but in exchange, 
Garry obtained serious compensation. 

The second critical moment of the game came when Deep 
Blue played 26. ... Bh7, allowing Garry to shut out its light square 
bishop for most of the rest of the game. We were not too happy 
with this move decision, nor were we happy about how Deep Blue 
was assessing the position as being in its favor. At best, the position 
was about equal. The penalty for bishops being shut out from the 
game was increased after this game. The alternative move, Bh5, 
appeared to be playable, but that would have been a different game 
with chances for both sides. Our main purpose in increasing the 
penalty was to encourage Deep Blue to try to activate the bishop 
whenever the opportunity presented itself. 

The rest of the game was not too remarkable. Deep Blue's 
move 40. ... Bc7, returning the pawn, was the last difficult move. I 
had replaced Joe as the operator in the middle of the game. Garry 
was not too happy during the last few moves in this game. Deep 



The Holy Grail 239 



8 

7 
6 
5 
4 
3 
2 
1 






yym, 



BHW. r!*,3&tn,v.. 



I! ■ &■ 



~ - "^ 



sua 
■ Bfe 









H HI 



* 






abcdefgh 

The position before 26. ... Bh7. 

Blue was in the slightly worse position, but it had set up an 
unbreakable fortress. For the last five moves, Deep Blue just shuf- 
fled its king back and forth, and there was nothing constructive 
that Garry could do. He knew it was a draw, I knew it was a draw, 
and the rest of the world knew it was a draw but after what hap- 



8 
7 
6 

5 
4 
3 
2 

1 






¥tSM 



m m 



m 






M 

mm £&£&** 



nMhrx """ 



m 



,,>.~-%*J««i* 



S Whw 



abcdefgh 

The position before 40. ... Bc7. 



240 Chapter 12 

pened in game two, Garry could not bring himself to offer a draw 
immediately. What if there was a win? 1 felt sorry for him when he 
deliberated long and hard before he was absolutely, positively sure 
and offered the draw. 

I was quite excited after this game. It was the first time we had 
ever played into a fortress draw in a serious game. 1 was not ready 
for Garry 's outburst when we got down to the auditorium. 

He first complained that Deep Blue spent the longest amount 
of time in all their games when it rejected the move Qb6 in game 
two. None of us were sure what Garry was saying but in any case, 
his recollection was wrong. Deep Blue rejected Qb6 twice in game 
two, at the time when it made 36. axb5 and 37. Be4. It spent about 
6.5 minutes at move 36. axb5, which was far shorter than the fif- 
teen minutes or so it spent on 35. Bxd6, the move right before 
axb5. Furthermore, I am almost certain that Deep Blue had spent 
more than fifteen minutes on other moves in earlier games against 
Garry. The second time Deep Blue rejected Qb6, it spent just over 
three minutes, which was its regular allocated time. It was not clear 
whether Garry was referring to 36. axb5 or 37. Be4. Some of his 
remarks during the match seemed to indicate that he was more 
concerned about 37. Be4, which, however, was just a positional 
move and was certainly not played at a slower pace. After the 
match, several computer chess programmers indicated that their 
programs would play 37. Be4 with slight changes in their evalua- 
tion function. Garry's next remark seemed to indicate that 36. 
axb5 surprised him. 

He then said, "I sacrificed two pawns and 1 had some chances. 
I don't think that these chances are enough." I had no idea what 
he was saying at the time. Now, looking back, it all made sense. He 
had seen the two-pawn sacrifice line that Yasser was talking about 
in the auditorium. That line did not quite work. At the time, none 
of us realized that the line Deep Blue had was an entirely different 
sacrifice. Garry then described that he tried out the lines on his 
computers, but none of them saw any way for Black to save the 
game in the two-pawn sacrifice line. 

He followed, "This machine missed, from a computer point of 
view, an elementary draw." When I heard this, 1 said to myself, 
"Huh?" The potential draw at the end of game two was hardly an 
elementary draw for a computer. Either he was getting very bad 
advice about what computers could do, or he genuinely did not 
understand chess computers. Both explanations were hard to 



The Holy Grail 241 

believe. He did have good computer chess people working with 
him, and had he tried to get Fritz to find the draw, he would have 
known that the draw was far from easy. If Deep Blue missed the 
draw, the draw obviously could not be an elementary one, unless 
we had a bug. Garry's remark just did not make sense. 

Joel was given the microphone and explained what had hap- 
pened in a straightforward way. I thought that he did a good job 
explaining everything in laymen's terms. 

Garry, however, continued to complain about how much time 
Deep Blue had spent on rejecting Qb6 (it spent only 6.5 minutes), 
how simple the draw was, and how the computer should not have 
been able to miss the elementary draw, and yet have been able to 
reject the Qb6 move. He also wanted to see move variations. 

At this point, Maurice Ashley said, "If I am reading you cor- 
rectly, Kasparov, or maybe I'm speaking out of turn. Do you think 
that there may have been some kind of human intervention dur- 
ing this game?" 

Garry said, "It reminds me of the famous goal which 
Maradona scored against England in 1986. He said it was the Hand 
of God." Maradona knocked the soccer ball into the net with his 
hand and the referee mistakenly allowed the goal to count. It 
seemed to me that Garry was apparently accusing us of cheating, 
although very obliquely. Every one of us was fuming. I was tempt- 
ed to grab the microphone, go to the Deep Blue model on the stage, 
and say, "Well, shall we open up this box and see whether Bobby 
(Fischer) is inside?" Over the years, Garry had been almost like a 
friend to me. He was as close as possible to being a friend without 
being one. The competition between the Deep Blue team and the 
Kasparov camp had been very good for both sides. Suddenly, I felt 
that I no longer wanted to have anything to do with him. 

Joel could not take it any more, "After the match, we'll be 
happy to share exactly what it saw in this case. But I think that it's 
definitely a mistake for Garry to give a position to Fritz or any 
other computer and say, 'This is computer behavior and this is 
what Deep Blue must be thinking or what Deep Blue would do.' I 
think he's seen from the games that he's played against Deep Blue 
that it is no ordinary computer and plays at an entirely different 
level from any computer that he's seen. So maybe he should come 
to grips with the fact that Deep Blue can do a lot of things that he 
did not think were possible." 

Joel could have been a little bit more diplomatic. Garry never 



242 Chapter 12 

did come to grips, not even after he was given Deep Blue's log for 
the two moves in question, after the match. I think part of the 
problem was that he asked for the log of the wrong moves. He 
should have asked for the log of move 35. Bxd6 in game two. Had 
he seen the 35. Bxd6 log, he would have realized that Deep Blue 
had already seen the three-pawn sacrifice when it finished its 10- 
ply search for Bxd6. 

The other part of Garry's problem was indeed that he did not 
know what Deep Blue was capable of. The decision to avoid Qb6 
on move 36 was mostly tactical, and involved mainly calculations. 
If he had really studied Deep Thought or Deep Thought II games, 
he would have realized that Deep Blue's predecessors were already 
capable of producing tactics way beyond the capabilities of other 
computers, even if the others had been scaled up to Deep Thought 
or Deep Thought ll's search speed. And Deep Blue was close to a 
thousand times faster than either Deep Thought or Deep Thought 
II. 

Last, but not least, the three-pawn sacrifice is quite hard, at 
least for other computers, and, it seems, for Garry as well. Pierre 
Nolot, a French chess journalist, once collected eleven chess posi- 
tions that he said were unsolvable by any computer in reasonable 
time. Deep Thought II solved several of them in tournament time, 
and maybe eight or nine of them within a day. Some of those posi- 
tions that Deep Thought II solved are probably still beyond the 
best commercial chess programs 8 , even when given a full year of 
time. The three-pawn sacrifice that could arise after 36. Qb6 might 
be in the same class of complexity. 

A Narrow Escape 

Before the start of game four, there was some serious doubt as to 
whether Garry would show up at all. His manager, Owen Williams, 
was on the phone for a while, and everyone was relieved when the 
report came that Garry was on his way. 

The Deep Blue team was mostly in the dark about what was 
happening until months later. The following is a reconstruction 
from interviewing the people involved. 

According to Ken Thompson, who was on the appeals com- 

8 This may change by the time you read this. A few amateur programs were doing 
well on the Nolot position after adding in singular extensions. Some commercial 
programmers may very well decide to adopt the idea as well. 



The Holy Grail 243 

mittee, the night after game two he received a request from Garry. 
He wanted him to go over the game two log and verify that every- 
thing was above board. Ken had seen the log in real time and was 
well aware of what happened, but wanted to see it again to refresh 
his memory before he reported back to Garry. 

Ken then asked C J for a printout of the log on the rest day 
and C J readily agreed to the request. Ken informed Garry that the 
request had been accepted. Later, C J promised to give Ken the 
game two log in question around noon the next day. Ken appar- 
ently told the other members of the appeals committee that the 
problem was taken care of. According to C J's recollection, some- 
one from the appeals committee, probably Monty Newborn, then 
told him that the problem was resolved. Unfortunately, this led to 
Ken not getting the log at the specified time, as C J thought that 
the problem no longer mattered. Somewhere around this time, 
someone in the Kasparov camp told C J that Garry wanted to see 
the log for himself. Owen Williams was later quoted as saying IBM 
had agreed to let Garry have access to all the game logs during the 
match but then reneged on the promise. As far as I know, no one 
from IBM had ever agreed that he could have access to the game 
logs while the match was still going on. One possibility was that 
people in the Kasparov camp might have misinterpreted what Ken 
told them. 

If it were up to the Deep Blue team, we would never have 
agreed to give Garry access to the game logs during the match. 
Deep Blue's game logs contained its "inner thoughts" on how the 
games would proceed, what it would have done given hypotheti- 
cal moves that did not show up in the games, what it thought 
Garry's best moves should be, and so on. If he had the logs, he 
would have a complete road map on how Deep Blue would behave, 
rightly or wrongly, in various circumstances. Even a rank beginner, 
armed with the game logs, a very good memory, and the help of a 
good team of Grandmasters, would be able to beat Deep Blue con- 
sistently by merely memorizing the winning lines. Garry, besides 
being possibly the strongest Grandmaster ever, has a phenomenal 
memory. Granting him direct access to Deep Blue's game logs 
would be equivalent to giving him Deep Blue's silicon head on a 
silver platter, with $700,000 on the side. 

Anyway, when Ken asked C J minutes before game three 
about the log, C J probably thought that he was asking whether 
Garry would be given the log, and gave a flat "no". Ken then told 



244 Chapter 12 

Garry that the request was turned down. Garry was most likely 
already not in the best frame of mind after discovering that the 
final position in game two could be a draw. The fact that he had 
White in game three and that game ended a draw probably did not 
help his mood either. No wonder he lashed out at the Deep Blue 
team after the game. This whole sequence of events did not make 
his outburst completely excusable in my mind, but did make it 
more understandable. 

By the morning of the day for game four, Ken was starting to 
get angry with C J, and said that he would not come to the opera- 
tions room unless the printout was produced. A phone call was 
placed by IBM security to Ken, who was in his office in New Jersey, 
that the printout would be available for him when he came back. 
Meanwhile, Owen was saying that Garry would not play. Garry 
later said that he never said he would not play. The whole thing 
was up in the air until close to game time, when it was reported 
that Garry was on his way. Ken finally got the log and informed 
Garry about its receipt ten minutes before game four. 

Now, back to the game itself. 

During game four, Yasser made the comment, "It looks to me 
like Deep Blue is in deep doo-doo", which filled the auditorium 
with laughter. Game four gave me the most anxious moment of 
the match, even though Deep Blue did manage to draw. It was also 
a game full of richness. Garry played a profound pawn sacrifice 
that Deep Blue had a hard time fully appreciating. Afterwards we 
also found the king safety evaluation bug that had shown up in 
game one. I am certain that if 1 look deeper into the game, I will 
find many new ideas to improve Deep Blue. 

Garry answered Deep Blue's 1. e4 with 1. ... c6 this time, the 
Caro Kann defense. Once again, he did not play his pet Sicilian 
defense, although by now this was no longer a surprise. We were 
fully expecting him to continue to play strange stuff. Garry's next 
move, 2. ... d6, however, had the commentators in the auditorium 
and the pressroom scratching their heads, "What do we call this 
thing?" 

Grandmaster Miguel Illescas was in the operations room for 
this game. I was also in the operations room for the early part. 
Miguel is one of the most cheerful people that I have ever met. He 
almost always found some good things to say about Deep Blue's 
moves. He was a great person to have around when we needed 
cheering up, and in this game we needed cheering up. 



The Holy Grail 245 



im 



p'- 






Mm/MiMk 



■'ill 



W//"V, 



I ■ 



ii fc ; 



i™ y - 






abcdefgh 

The position after 2. ... d6. 

Miguel was quite optimistic about Deep Blue's chances early 
on. By move twenty, the machine seemed to have all the opening 
problems solved and was ready to start some real action. Right at 
this point, Garry unleashed an amazing pawn sacrifice 20.... e5. 
When Miguel saw this move, he instinctively called it a good sac- 



8 


§f#jplli m 


7 


■Anii ■ 


6 


u mmmm 


5 


111 •.*■ HI 


4 


■ i aaii 


3 


■ ■ Kfl 


2 


AH&Bwa ■ 


1 


■*■£« ■ 



abcdefgh 

The position before Kasparov's 20. . .. e5!. 



246 Chapter 12 



rifice. Deep Blue would not play the move but assessed the sacri- 
fice as sound. 

Next, Deep Blue played some moves that were quite strange 
to human eyes. 26. b5 was both panned and praised by different 
commentators. Garry himself spoke highly of the move. We did 
not like the move, and it triggered an evaluation software bug hunt 
the next day. Joe was the one who finally found and fixed the bug 
relating to this move and the corresponding move from game one. 
Deep Blue's strange moves right before playing b5 appeared to be 
caused by the same bug. 



8 
7 
6 
5 
4 
3 
2 
1 



m f 



NSifcr 



rini 



v^ 



■; 



fU 









m 



fe : 



m. ■< 



^. 



2L 



abcdefgh 

The position before Deep Blue's 26. ... b5. 

As the position started to clarify after Deep Blue's strange b5, 
Miguel was having second thoughts. The position continued to 
deteriorate, and even Miguel was getting worried. Deep Blue was 
starting to evaluate the position as better for Garry. The day had 
started brightly for us, with two Whites for three games, but now 
things were looking bad. 

At this juncture, we had a surprise guest in the operations 
room. IBM Chairman and CEO, Lou Gerstner, had decided to pay 
the team a visit. New York Times reporter Bruce Weber happened to 
be in the room. Lou's dramatic comment on the match was print- 
ed in the New York Times the next day: "1 just think we should look 
at this as a chess match between the world's greatest chess player 
and ...", a short pause, "Garry Kasparov." 



The Holy Grail 247 

After this, none of us was willing to spoil the moment and tell 
Lou that Deep Blue might be in trouble. Lou stayed for about five 
minutes and then left. 

When Miguel had time to look at the board position again, he 
became quite pessimistic, the most pessimistic that I ever saw him 
in the entire match. He thought that Deep Blue would not be able 
to survive the ending, which looked unfavorable. Deep Blue's 
assessment of the position, however, never got below half a pawn 
down. Murray was operating Deep Blue at the time. If Miguel was 
so unhappy about the position, Murray must have been feeling 
even worse. 

On Garry's move forty-three, Deep Blue self -terminated. Before 
the match, Joe added in a piece of code that monitored the efficien- 
cy of the parallel search and terminated the program itself if the 
efficiency dropped too low 9 . Joe's idea was that a fresh start would 
be preferable in this case. This instance was one of the two times dur- 
ing the entire rematch that the self-termination code got executed. 
In game three, Deep Blue also self -terminated, but it was very early 
in the game, right after Deep Blue's move eight, and while Garry was 
still thinking about his move. We had to wait until he made his 
move before restarting the program on our own time in that game. 

I took over as the operator after Deep Blue's game four self -ter- 
mination. Garry asked briefly what happened and was satisfied 
when I told him that the program had to be restarted. Garry's 
move forty-three happened to give Deep Blue only one legal reply, 
so he knew precisely what the machine would play when it came 
back. Our problem gave him free time to think on the position, 
meanwhile, Deep Blue's clock was ticking. After the rematch, we 
had heard that Garry was upset that Deep Blue "crashed" numer- 
ous times and that he may have suggested that the "crashes" were 
distracting and suspicious. Personally, I would have preferred that 
Joe did not use a mechanism like the self-termination code, but 1 
was not in Joe's shoes. However, I strongly doubted that Garry was 
really distracted by the two instances of self-termination and there 
was certainly nothing to be suspicious about, especially in the case 
of game four. You cannot violate the chess rules and play an illegal 
move. 



9 For the readers who are also computer scientists, the efficiency could drop too 
low when there is a deadlock in the parallel system. Also, things can go wrong 
for other unknown reasons. 



248 Chapter 12 

I was expecting a disaster when I left the operations room, but 
by the time Deep Blue was restarted, the position looked much bet- 
ter. Previously, I was just hearing Miguel saying that it was 
hopeless. Now I was examining Deep Blue's lines for myself, and 
those coming up seemed to point to a draw. The position on the 
board was a rook ending. The chess chip contains endgame ROMs 
(Read Only Memory) that encapsulate a lot of the relevant rook 
ending knowledge. Deep Blue also had access to the five-piece rook 
ending database on line, although in this game the database did 
not matter. After Deep Blue's move fifty-six, the position had 
reached a clear draw. Garry opened up his hands and said, before 
making any move, "Draw?" This draw offer was unorthodox. 
Normally, one offers a draw right after making a move. I immedi- 
ately accepted the offer. 

Did Garry have a win somewhere in the endgame phase? It is 
unclear. Many analysts claimed that they found winning lines, but 
on closer examination none of them works. 

The match was all tied up with two points each after four 
games. 

After a Drink or Two 

Game four was followed by two rest days. In the Philadelphia 
match, there was only one rest day between game four and game 
five but during the negotiation for the rematch, Garry explicitly 
requested the extra rest day. The extra day probably cost IBM at 
least $100,000 more, in leases alone, for the venue and the equip- 
ment. 

Of course, rest days were not really "rest" days for the Deep 
Blue team. The three of us, along with the Grandmasters, were hard 
at work trying to resolve problems that showed up in the match 
games, or finding potential problems in practice games or test posi- 
tions. The critical evaluation bugs got fixed up by Joe. One of the 
match games triggered Murray to make a conceptual change to the 
evaluation function. We were learning from Garry himself. I ran 
Deep Blue Jr through a vast set of test positions and located a few 
bugs, which Joe promptly fixed. The Grandmasters worked on the 
openings, although it was clear by now that Deep Blue's opening 
preparations probably would not be tested. When we completed 
the bug fixes, the Grandmasters played a few test games against 
Deep Blue Jr to make sure that the modified program was indeed an 
improvement. This went on for most of the two rest days. 



The Holy Grail 249 

In game five, Garry reverted to the basic setup he played in 
game one, but with a slight deviation. On move four, Deep Blue 
decided to trade off its bishop for his knight, giving him the 
advantage of the bishop pair. But in exchange for the bishop pair, 
Deep Blue would have better piece development. In the operations 
room, we could see this from Deep Blue's main line. Giving up the 
bishop was a perfectly reasonable choice. None of us on the Deep 
Blue team thought much about the move. It was just a move. But 
we also knew that Garry might have problems with it. Some com- 
ments from Kasparov's camp indicated that they had a very 
simplistic view of Deep Blue's evaluation function. After game one, 
the Kasparov camp (as did Yasser Seirawan) apparently thought 
Deep Blue valued the bishop much higher than the knight, and it 
would go through serious contortions to preserve it. Garry's remark 
at the post-match press conference indicated that he was indeed 
surprised by Deep Blue's decision to trade off the bishop for the 
knight. I wondered whether he considered the possibility that 
Deep Blue might, indeed, be far more sophisticated than he gave it 
credit for. 

Several members of the press apparently had taken up Garry's 
simplistic view about how Deep Blue evaluated a position. During 
this game, I went to the pressroom to give a few interviews, and 
was surprised by repeated questions about whether we had adjust- 




-j y i j , . j 



iii 




MAS if/* 1 



z 



Iht?}'/, « — ' '/A.... 



abcdefgh 

The position before 11. ... h5. 



250 Chapter 12 

ed the relative values of bishop and knight before the game. The 
reporters meant well, but the questions felt like being asked, "You 
used your left hand to pick up the book just now. Did you change 
the relative values of your left hand and your right hand?" 

After Garry's eleventh move, Deep Blue had a significant lead 
in piece development. At this point, Deep Blue produced the move 
h5, pushing its h-file pawn two squares. Garry started to jot down 
the move on his score sheet, then he took a double take and 
glanced at Murray, who was operating for Deep Blue. Murray gave 
Garry a slight nod, indicating that Deep Blue did play the move. 
Garry was surprised. So were all of us in the operations room. Joel's 
first reaction was, "I don't like the move. But I can't say what is 
wrong with it." In the auditorium, the immediate response from 
Maurice Ashley was "No!" Yasser: "Garry is a happy camper! Who's 
been programming this machine?" Maurice: "I mean we've seen 
some strange moves and we've seen some strange moves. That's a 
strange move!" Miguel apparently did not quite like the move at 
first either but after a moment's thought, he said, "Well, maybe 
one or two of the top ten players would play it, but only after a 
drink or two." I never did ask Miguel what he meant. But I was 
interpreting his comment the following way. The move h5 was sig- 
naling to Garry, "If you castle king side, I will attack you." 
Normally, even a top-ten player would not dare to be so aggressive. 
Deep Blue obviously had no idea that it was playing Garry 
Kasparov. Garry's post match comment indicated that he was quite 
shocked by the move, "No computer plays h5!" It was quite inter- 
esting that, during the game, the reaction in the pressroom was 
that h5 was a computer move, with some people inserting "stupid" 
before "computer". Analysts of the game later praised the move as 
a sophisticated positional idea. 

Two years later, the move h5 was at the center of a controver- 
sy over the Internet. When I first heard Garry's post-match 
comment that "no computer plays h5", I had no idea whether 
Garry was praising or accusing us. Right after the game itself, Clara 
Kasparova asked her son whether there was anything amiss. He 
replied that it was a typical computer game so there was certainly 
the possibility that he was praising us. In 1999, however, when 
interviewed by a chess journalist, he seemed to imply that his 
statement was an accusation. The chess journalist posted on the 
Internet what we understood to be Garry's assertion that no com- 
puter played h5, and therefore we had cheated. Within hours, 



The Holy Grail 251 

someone on the Internet pointed out that at least one commercial 
chess program also played h5. The chess journalist happened to 
know that Garry had the very same commercial chess program on 
his own laptop a few feet away during the interview. So he not 
only wrongly accused us, but it appears he did not even bother to 
check whether any of the chess programs he owned played h5. I 
don't know why the commercial chess program played h5. It could 
be good programming, bad programming or just dumb luck. 

When I saw the move h5 from Deep Blue, I knew precisely 
what hardware evaluation features prompted the move. During the 
last two months of chip design before the rematch, I added drastic 
changes to the hardware for king safety evaluation. Before the king 
castles, the hardware computes three sets of king safety evalua- 
tions, one for kingside castling, one for queenside castling, and 
one for staying in the center. The real king safety evaluation is the 
weighted linear combination of the three, with the weighting 
based on the relative ranking of the three, and difficulty of making 
the castling moves. Normal chess programs have a far simpler king 
safety evaluation. To avoid problems with king safety, chess pro- 
grammers usually actively discourage their programs from playing 
moves like h5, because it damages the potential king safety for 
castling kingside, even though in reality, the programs might be 
able to castle queenside safely. Since Deep Blue could tell whether 
it could castle queenside safely on its own, we just allowed it to do 
whatever it wanted. In the game position, Deep Blue could always 
castle queenside safely, and therefore move h5 was perfectly 
playable from its point of view. 

After Deep Blue's move twenty-two, it assessed that it had a 
slight edge. I was getting optimistic, "Could we possibly win this 
game?" Just when I was getting excited, 1 was called up to the 
pressroom to give interviews. Garry apparently thought he was in 
trouble as well. By the time I got back to the operations room, 
however, he had skillfully tiptoed around any potential problems. 
Around this time, Deep Blue played a move that led to some posi- 
tional problems but Garry did not play the best continuation. 

By move forty, Deep Blue assessed the position as being even. 
Garry later stated that he saw at this point how the game would 
end in nine more moves. The commentators on the auditorium 
stage did not see the game finish until the very end. The game fin- 
ish was quite pretty. At first glance, Garry had a passed pawn that 
seemed very difficult to stop. Deep Blue, instead of trying to stop 



252 Chapter 12 



8 

7 

6 

5 
4 

3 
2 

1 




Up — § 

p g p ■"■" 

• II 10 in 









m?, m 



mz 






abcdefgh 

The final position in game 5. White to move. 

the passed pawn, threatened to checkmate. After Garry fended off 
the mating threat, Deep Blue just marched its king forward, ignor- 
ing the passed pawn. The final game position had Garry ready to 
promote his passed pawn to a queen, but it was all to no avail. 
Deep Blue was about to initiate a drawing sequence based on rep- 
etition checks. 

Garry stayed at the chess table after the draw was agreed, and 
began talking to his mother in Russian. Soon the game room was 
filled with people. Several of Garry's people were there as well as 
Ken and C J. In the operations room, we were at a loss as to what 
was happening. We were told not to go over to the game room 
yet. Word soon came that Garry wanted all the game logs to be 
printed, sealed, and placed in the possession of the match arbiter, 
Carol Jarecki. The day before, he had made a formal request that 
all the game logs be printed and placed in sealed envelopes. The 
appeals committee had voted against requiring IBM to do so, but 
had also suggested that IBM should consider doing it. Garry was 
told about the decision of the committee. Just prior to the game, 
an internal IBM decision was made that we would honor his 
request, reversing an earlier decision. So the problem was quickly 
resolved, although not without some hiccups. The computer 
printer had a paper jam that took a while to fix, and there was not 
enough paper nor enough time to print all the game logs. Only 



The Holy Grail 253 

the log for game five was printed. The next day, two complete sets 
of game logs were printed and given to Carol Jarecki and Ken 
Thompson. 

After Carol got the seventy-two pages of game five's log safe- 
ly in her briefcase, she went up to the pressroom and briefed the 
press about what happened. Carol had the distinction of also being 
the match arbiter for the 1995 Kasparov vs. Anand World 
Championship match. Concluding her briefing, Carol made a 
comparison between the two matches, "I'll tell you one thing. This 
match is a lot more exciting than the World Championship." 

She did not mean just the games. 

A $300,000 Gamble 

After five games, the match was tied at 2.5 points each. Deep Blue 
had White in the last game. Unless Garry could somehow manage 
to win the last game from the Black side, history would be made. 
Deep Blue would be either the first computer to tie the World 
Chess Champion in a regulation match, or better still, the first 
computer to beat the World Chess Champion in a regulation 
match. 

The scalpers were out in full force, reselling the $25 tickets for 
as much as $500. If you had bought one of the $500 tickets, you 
would have paid about $10 for every minute of game time, as game 
six lasted less than an hour. 

Personally, before game six, I felt a sense of calm. Murray, Joe, 
and I had worked long and hard to reach this stage. The match was 
tied. Deep Blue had White, so our chance of at least drawing the 
match was quite good. We had, however, no real control over what 
would happen next. "What will be will be," I told myself. Yes, my 
twelve-year endeavor might finally be over, but I could only watch 
at this point. The only thing I wanted was just a good — preferably 
great — game by both sides, and that the historians would say about 
the match, "Neither side deserved to lose." I did not care whether 
the anti-computer Garry or the normal Garry showed up. I just 
wanted a great fight as befitting the potentially historic occasion 
after such a closely contested match. 

Most of the IBMers were also in great spirits, with the excep- 
tion of the PR people. They were bracing for a firestorm should 
Deep Blue draw or win the match. Garry had been dissatisfied 
when the match was still tied. What might he say if he lost a 
match for the first time in his life? Yes, you read it right. Garry had 



254 Chapter 12 

never lost a single chess match in his professional life before the 
1997 rematch. He had come close to losing to Anatoly Karpov in 
their first world title match, but that match was cancelled when 
Anatoly's health seemed to be failing. Some were concerned that 
Garry would react angrily to losing a match. I was surprised at this 
notion. The rematch was in many ways bigger than even the 
World Chess Championship matches. The PR people were making 
preparations for a volcanic eruption. The IBM team was asked to 
dress up for the occasion, and also asked specifically not to smile 
during the closing ceremony, especially if Deep Blue won the 
match. The PR people did not want to fan the fire should a 
firestorm erupt. 

The atmosphere in the auditorium before the start of the 
game was festive. Video clips from the TV shows 'Late Night with 
David Lettermari and 'The Tonight Show with Jay Leno' about the 
rematch were shown, to laughter and applause from the audience. 

Joe was the operator for the last game, with Murray as the 
backup. I had decided positively and definitively that I would not 
work on this particular day after twelve years. This was going to be 
my day as just a spectator. 

The game began the same way as game four, with a Caro Kann 
defense. Instead of playing a weird sideline as he did in Game four, 
this time Garry played the main line. Deep Blue also played the 
main line for White; in fact, the main line that Garry as White 
used against Anatoly. Garry was in a familiar territory, albeit from 
the opposite side of the chessboard. 

On Garry's seventh move, after spending nearly two minutes, 
he played 7. ... h6. Deep Blue instantly replied Nxe6, giving up a 
knight for a pawn. Garry acted a little bit surprised, but then 
played the next few moves very fast, as did Deep Blue. Both sides 
were still in book. 

In the auditorium, Yasser was saying that Garry had blun- 
dered and transposed the move. He also said that Garry was in 
terror and distress. None of us in the operations room, including 
Ken, believed this. It was inconceivable that Garry, with his leg- 
endary memory, could have forgotten an opening line that he had 
played many times, albeit usually from the White side. Garry's 
expression was also more one of surprise rather than terror and dis- 
tress. Those emotions would come later. 

So what did happen? An International Master in Kasparov's 
camp was quoted the next day in the newspaper that the 7. ... h6 



The Holy Grail 255 



8 

7 

6 

5 
4 
3 
2 

1 




abcdefgh 

The position after Kasparov's 7. ... h6. 

move was one agreed upon earlier. Garry himself stated months 
later in an interview that he regretted the decision to play 7. ... h6. 
So the move was never a slip of the finger as it was characterized 
in news articles immediately following the match. It was played by 
design. 

So why did Garry play the move? Black's position was gener- 
ally considered difficult at best after the knight sacrifice. This is 
true in games played by human players, but it is not true in games 
played between computers and human players. The commercial 
chess programs apparently had serious problems avoiding losing 
the game as White! Several Grandmasters had tried playing the 
Black side of the game positions against the top commercial chess 
programs and were able to win every single game. This fact did not 
come out until well after the match, but by then the media had 
moved on to other stories. What Garry played in game six on move 
seven was a very risky anti-computer chess move. 

How did Garry get the idea to play 7. ... h6? In the post-match 
news conference, he seemed to blame his helpers. There were other 
possible sources. Several of the top commercial chess programs at 
the time were explicitly prohibited in their opening books from 
playing the knight sacrifice that Deep Blue played. So, apparently, 
lots of the commercial chess programmers knew that their pro- 
grams could not play the sacrificial line. There were some rumors 



256 Chapter 12 

during the match that Garry thought Deep Blue was using the 
opening book of one of the commercial chess programs, so it was 
conceivable that he was expecting Deep Blue to retreat its knight as 
the other programs did. This would have given Garry a much bet- 
ter position than he could have achieved in the normal lines. In the 
"unlikely" event that Deep Blue did play the knight sacrifice, well, 
there was this nice article written by Grandmaster Timoshchenko, 
one of Garry's former seconds, in the March 1997 issue of the 
International Computer Chess Association Journal. Timoshchenko 
played exactly the same line against none other than Fritz, the 
chess program distributed by Garry's computer chess consultant, 
Frederic Friedel. Timoshchenko gave the verdict, "After the knight 
sacrifice, Black has enough possibilities for defense." 

So there were lots of potential reasons to select the move 7. ... 
h6. First, Deep Blue might just retreat the knight. In this case, 
Garry would have equalized very easily. Second, Deep Blue might 
not "understand" how to attack. Then, Garry would have won the 
game and become the toast of the human race. (Although Deep 
Blue, in my opinion, might be the more important human 
achievement when all was said and done.) Third — and this one a 
stretch — if Deep Blue did "understand" the position and beat him, 
a bystander could always say that the last game did not count and, 
given that game two could have been drawn (as game one could 
have), he was the actual winner of the match. Believe it or not, 
there were actually people who used this argument to suggest that 
Deep Blue was the loser of the match. It would be way too cynical 
to believe that Garry had planned this third scenario. 

Deep Blue was out of its opening book on move 11. Bf4. Its 
first evaluation of the position was that it was close to a pawn up, 
even though it was really down about two pawns in material over 
the board. Deep Blue saw close to three pawns worth of positional 
compensation. At about this time, Miguel popped in and asked 
what was Deep Blue's evaluation of the position. After we told him, 
he smiled broadly. All of us knew that the match was effectively 
over. Once Deep Blue had the initiative, and knew that it had the 
initiative, there was no stopping it. The position on the playing 
board had actually shown up in our lab about a month earlier 
when the other Grandmasters were working on Deep Blue's open- 
ing book. At the time, the Grandmasters did not bother to play out 
the position either; they just looked at Deep Blue's evaluation of 
the position and moved on to other work. They all had been on 



The Holy Grail 257 



8 

7 
6 
5 
4 
3 
2 
1 



ww WW 




abcdefgh 

The position after Deep Blue's 11. ... Bf4. 

the receiving end in similar positions. Miguel was not in the 
United States at the time of the discussion, so he had not yet seen 
Deep Blue's evaluation of the position. Maybe Black could draw 
the position, but probably no human player could hold the posi- 
tion against Deep Blue. It appeared that Garry's $300,000 gamble 
would backfire on him. 

Garry's demeanor at the board quickly changed from sur- 
prised to distressed, as Deep Blue spurned a quick gain of small 
material and continued its attack. Near the end of the game, Garry 
started talking to Clara in Russian. It is against the rules of chess to 
talk to spectators while a game is still going on. It was not clear 
what he said to his mother. Perhaps he told Clara that there was 
no sense in playing on any more. After Joe played 19. c4 for Deep 
Blue, Garry looked at Joe in a questioning manner. Joe nodded 
slightly, confirming what he was suspecting. Deep Blue was win- 
ning major material. Garry then promptly resigned. The video 
segment of his resignation would be played many times, around 
the world, in the next few days. After his resignation, Garry shook 
hands with Joe and just walked straight out of the room with both 
arms up in the air the entire time he was walking out. It was a sur- 
real scene. The game had lasted less than an hour. 

At the moment of Garry's resignation, I suddenly felt very 
tired. The twelve years of work were finally over. I should have 



258 Chapter 12 

been exulting, but I was feeling empty inside. The game felt too 
easy, although in hindsight it wasn't. Without our hard work the 
year before, Garry might have won the game. A part of me also felt 
robbed. I am not a chess player, but the chess player in me was def- 
initely disappointed. Win or lose, I had wanted the last game to be 
a real fight. I wanted the win to be another great game like game 
two, except without the final mistake. If it had to be another loss, 
I wanted it to be another hard-fought loss like game one, prefer- 
ably without the bad moves from Deep Blue. 

So, almost fifty years after Claude Shannon made the propos- 
al on how to program a computer to play chess, the quest for the 
holy grail was finally over. It did not end with the greatest chess 
game ever, but how the match ended was not up to us. 



8 
7 
6 
5 
4 
3 
2 
1 



■ 






i! mm m 

W^ a M^ y ^M '^i?/- 
Wm w mm imv/ ym& 

, t Will A !// 



■ AH 






P 

m. 4m. 



7>777i7. 



m 



m 



'^'■-mi^ 



rf 



abcdefgh 

The final position of the match. 



Epilogue 



CHAPTER 13 



Life after Chess 



The Press Conference 

At the post-match press conference, Garry characterized the match 
as "lost by the world champion". He then lashed out at IBM and 
he seemed, to me at least, to be hinting at human intervention in 
game two. He also once again wrongly characterized the potential 
draw by repetition checks in game two as "simple". There was also 
some bravado. "I personally guarantee you, I'll tear it to pieces 
without question." "It will not help, because we know how the 
machines play." "I think the machine has too many weaknesses." 
One of the more surprising statements was "But it's not yet ready, 
in my opinion, to win a big contest." I don't know what he had 
been thinking, but in all media metrics, the rematch was a far big- 
ger event than any other that he had played in. The number of 
"impressions" registered for the rematch alone, not counting post- 
match publicity, was 5 to 7 billion. This was probably 100 to 1000 
times the number of impressions registered for any of his World 
Championship matches. If the rematch was not a big contest, I 
don't know what is. 

On game six, Garry said, "Today's game doesn't even count as 
a game because probably it has been published before elsewhere. 
When the computer takes on e6 you can resign. I was not in the 
mood of playing at all, I have to tell you." I shook my head in my 
mind when I heard this. At the time, I was just appalled at the total 
absence of sportsmanship in the comment. I was not as certain as 
I am now that Garry had just made a $300,000 gamble and lost. 

Going back to game two, Garry again stressed, "Forget today's 
game. I mean Deep Blue hasn't won a single game out of the five 
because again in game two, I resigned when I could have forced a 

261 



262 Chapter 13 




Kasparov speaking in the press conference 
after game 3. 

From left to right: Mike Vaivo, Carry Kasparov, Feng-hsiung Hsu, Cj Tan, 

Joel Benjamin, Murray Campbell, joe Hoane, and jerry Brody. 

Kasparov was about to begin his public accusations. 

draw." He conveniently forgot to mention that he did not see the 
drawing continuations. By the same argument, we could have also 
said that he did not win a single game out of the five, as Deep Blue 
played a wrong move to lose game one, just as he made a wrong 
move, resignation, to lose game two. 

On future matches, Garry wanted, "One condition, IBM as 
player, not a sponsor." I would not mind for myself getting the 
royal treatment that he was getting. To the best of my knowledge, 
the part of IBM that was the sponsor was attending to his every 
need, within reason. He was not getting the game logs from the 
IBM player while the match was going on, but this was simply a 
request for an unfair advantage. The match would have been a 
sham if he were given free access. 

This was followed by another big surprise. I may have misin- 
terpreted what Kasparov was saying, but it appeared he was 
accusing the Deep Blue team of cheating in game one, the one 
game we lost: "You know, it [the anti-computer strategy] was work- 
ing, but suddenly it stopped working — suddenly Deep Blue found 
a way just to break the pawn chains and start a confrontation in a 
very, very convenient situation." 



Life after Chess 263 

Garry eventually calmed down a little bit: "But obviously it's 
a historical achievement that the machine was even able to play 
on such a level with the World Champion." 

More on future matches: "I think eight or ten games, a nor- 
mal match." Later, he added, "I think we should play every second 
day. You should give a human being time to rest. You know, twen- 
ty days, ten games, a proper match." 

Personally, I have some doubt whether a sponsor could be 
found for a match of the duration that he suggested. Excluding the 
prize fund, the extra length of the match would cost the sponsor 
in excess of one million dollars. The extra days could significantly 
reduce the media interest, especially with half being just rest days. 
There are good reasons why the NBA Championships and profes- 
sional baseball's World Series are seven games each. It is difficult to 
keep the full attention of the media on a sports event for much 
longer than a week. His proposal would lead to an event lasting 
nearly three weeks and with nothing happening on half of the 
days. 

The press conference lasted about thirty minutes. During the 
entire time, the Deep Blue team sat and listened to his monologue. 
We were asked not to smile to express our joy, nor could we voice 
our outrage at his implications. We had just accomplished what we 
had dreamt and worked for years: twelve years for me, about nine 
years for Murray and seven for Joe. And yet, we had to sit there and 
listen to these charges. It was one of the worst experiences of my 
life at a time that should have been my most joyous moment. 

Afterwards, Joe said something that I remember to this day, 
"You know, he [Garry] ruined the joy of it." 

The Old Gang 

Two months later, we were at a more joyous occasion. On 29 July 
1997, in Providence, Rhode Island, at the AAAI national confer- 
ence, Murray, Joe and I received the Fredkin Prize for the first team 
to build or program a chess machine to defeat the World Chess 
Champion in a match. The $100,000 Fredkin Prize, which was 
one-seventh of the winner's prize for the rematch, was split even- 
ly three ways, excluding the extra cent that was kept by the trustee 
of the Prize. 

Right after the rematch, we did not believe that we would 
receive the Fredkin Prize. While satisfying the original intent of the 
Prize, the rematch did not conform to the Prize Committee's 



264 Chapter 13 

match conditions set a few years back, in particular the length of 
the match and the size of the prize fund. Shortly after Murray and 
I joined IBM, Dr Berliner, the chairman of the Fredkin Prize 
Committee, and our former royal opposition at Carnegie Mellon, 
issued a public statement specifying the required match conditions 
for claiming the Prize. At the end of his statement, Dr Berliner 
admitted that it was perfectly possible that his match conditions 
might not be to the liking of the World Champion and the match 
sponsor. For the Deep Blue matches, as far as I know, no one at IBM 
even bothered to look at Dr Berliner's list of match conditions 
when setting up the two matches. 

The Fredkin Prize Committee did not make any announce- 
ment at the conclusion of the rematch; it appeared that the 
committee considered Deep Blue's match win inadequate. 
Winning the rematch was rewarding enough for us, so we did not 
pursue the possibility of claiming the Prize. It came as a pleasant 
surprise that, about a month after the rematch, Professor Raj Reddy 
informed us for AAAI that we had won it. What changed the mind 
of the committee? I have no idea. Maybe they realized that, for all 
intents and purposes to the world at large, the quest was over. 

After the awarding of the Prize, 1 gave a presentation to the 
assembly on the contribution of Deep Blue. I was moved when 1 
received a standing ovation before 1 started the presentation. In 
New York, after game five, one of the better games by Deep Blue, 
the partisan crowd booed us. This time, I was among friends, 
friends who understood that the match was never really "man ver- 
sus machine", but rather "man as a performer versus man as a 
toolmaker". Whatever the outcome of the match, when we 
cheered the winner, we were cheering for a unique human 
achievement. When Garry won in 1996, the Deep Blue team 
cheered and applauded his outstanding personal performance. 
When Deep Blue won in 1997, society at large, not including Garry 
Kasparov, finally recognized Deep Blue for what it was, namely, the 
advancement of a powerful tool created by human beings. 

Also present at the AAAI conference were several of our old 
friends from Deep Thought days, accepting the Allen Newell 
Award for Research Excellence for our joint work with Deep 
Thought. Thomas Anantharaman had left the Wall Street firm, and 
was working as an assistant professor at New York University, 
doing research related to the human genome project. Andreas 
Nowatzyk had left Sun Microsystems and joined Digital 



Life after Chess 265 

Equipment Corp. When we had dinner together after the ceremo- 
ny, Andreas mentioned the ambitious, personally-funded project 
he was working on in his spare time. He wanted to create a com- 
plete 3-D map of a frozen mouse down to 0.2-micron resolution, a 
resolution good enough to map most of the neural connections of 
the mouse's entire nervous system, including the brain. Mike 
Browne did not show up for the award ceremony. He had also left 
the Wall Street firm, had since worked at Sun Microsystems, and 
then joined Yago Systems, a successful startup working on smart 
switches for network connections. 

Moving Pieces 

What happened with the main characters after the rematch? 

In early news articles, Deep Blue, the victor, was reported to 
have been disassembled and shipped to IBM customers, without 
the chess chips. Its death, however, was exaggerated. Due to its his- 
torical significance, as of this writing, Deep Blue is still alive and 
kicking in IBM Research. Early internal reports stated that only the 
outer frames of Deep Blue stayed in IBM. It turns out that only the 
RS/6000 nodes went back to the factory test floor, and have since 
been put back into the original frames. Deep Blue no longer has all 
the chess cards that it used in the rematch. Two cards went to IBM 
headquarters in Armonk, where a Deep Blue Jr was set up to give 
demos to visitors. Most of the cards are still in IBM Research. The 
1997 Deep Blue now has half of its original chess cards. Some of 
the cards are placed in the old RS/6000 SI J used in the 1996 match. 
A few are placed in standalone workstations. The rest are on 
shelves. 

The Deep Blue RS/6000 SP is being used mainly for research 
projects unrelated to chess. It can still play chess — it serves as a web 
site that can play against up to one-thousand chess players simul- 
taneously. The web-based program is, however, much weaker than 
even "Pico" Deep Blue Jr, the single chip and reduced speed ver- 
sion of Deep Blue Jr. The web-based program uses a single Deep 
Blue chess chip, at one second per move, including setup over- 
head. The actual computation time is probably less than half a 
second per move. The very short computation time means that the 
web-based program is comparatively weak tactically, more so than 
typical chess programs would be with shortened computation 
time. Most of the search extensions that are partially responsible 
for the tactical strength of Deep Blue Jr are effectively inactive, 



266 Chapter 13 

given the short computation time. This program also uses a sim- 
plified evaluation function, does not keep track of move repetition 
and does not "think" on the opponent's time. It is probably no 
stronger than about 2000 rating on the USCF scale. This is fine, 
normally, as the opponents for the web-based program are usually 
IBM customers, and it is perfectly acceptable to have happy IBM 
customers winning a few games against the program. In 1999, the 
International Computer Chess Association asked IBM to provide 
the web-based program to the spectators in a computer chess 
event. A commercial chess programmer took advantage of the poor 
defenseless web-based program by playing it against his own pro- 
gram, giving his own program a hefty time advantage of the order 
of 50 to 1. (To be fair to the programmer in question, he probably 
did not know what the actual time advantage was when he played 
the games.) The commercial chess programmer then posted the 
games that his program won on his own web site as evidence of the 
"superiority" of his program, which he is perfectly happy to sell to 
you. 

As of this writing, Garry Kasparov has not played serious 
games publicly against any computer since the Deep Blue match- 
es. His performance against human players has its dramatic ups 
and downs. 

In a tournament right after the rematch, Garry punished the 
first human chess players he played and won the tournament fair- 
ly easily. For a short while after the rematch, his chess results were 
somewhat spotty; his match loss against Deep Blue probably pro- 
vided an initial psychological boost to his human opponents. 
Then he had a phenomenal run of great chess performances, with 
his chess rating going up to the middle 2800s, well above any 
other human chess player. But the cracks in his armor that showed 
up in the Deep Blue matches eventually came back to haunt him. 

By the year 2000, Garry had not defended his World 
Championship for over five years. The normal World 
Championship cycle was three years, and the recent trend is for 
even shorter cycles. In recent years, FIDE, the international chess 
federation, has held its own World Championship tournament 
once per year. While no one doubted that Garry was the strongest 
(human) chess player, there were people who considered the World 
Championship title vacant until he played a serious challenger. He 
had had three challengers since his 1995 title match against 
Vishwanathan Anand. The first was Alexei Shirov who won the 



Life after Chess 267 

right to challenge him by finishing second at the Linares tourna- 
ment, ahead of Garry but behind Anand, whose written contract 
with FIDE prevented him from playing in 1998 for any non-FIDE 
world title. The match between Garry and Alexei never took place. 
In 1999, Garry designated Anand as the new challenger without a 
normal qualifying event, bypassing Alexei. Anand was ranked 
number two in the world, so there was some public support for the 
new choice, even though many people did not like the way that 
Alexei was treated nor the arbitrary way that the challenger was 
chosen. The scheduled new match with Anand also never took 
place. Early in 2000, Garry designated Vladimir Kramnik as the 
new challenger, replacing Anand. Anand had a dispute with the 
match organizer over contract conditions that he considered unac- 
ceptable, and was replaced as a result. The Kasparov vs. Kramnik 
match took place in October 2000, with the surprising result of 
Vladimir ending Garry's fifteen-year reign by beating him in fif- 
teen games with the score of two wins and thirteen draws. Anand 
did not sit idle and became the FIDE World Champion himself. As 
of writing, Vladimir is the Braingames (the match organizer back 
in 2000)/Einstein (the acquirer of match rights from Braingames in 
2002) World Champion, 18-year-old Ruslan Ponomariov is the sur- 
prise 2002 FIDE World Champion, and Garry is without a world 
title but still the top ranked player, with Vladimir closing in as the 
second ranked player in the world. 

The Grandmasters who worked on Deep Blue have generally 
had good years since the rematch. None of the opening lines pre- 
pared for Deep Blue were really ever used, and the Grandmasters 
had very powerful secret opening weapons as a result of the coop- 
eration with Deep Blue. Joel Benjamin and Nick DeFirmian became 
the 1997 and 1998 US Chess Champions respectively. Of course, 
they are strong Grandmasters in their own right, and had good 
chances to win the title with or without their work with Deep Blue. 
Miguel Illescas has the distinction of being Vladimir's match sec- 
ond in his successful title bid over Garry, and thus becoming the 
only person to be involved with the opposing teams in both of 
Garry's match losses. 

The Deep Blue project had wound down in IBM. Murray and 
Joe moved on to financial modeling and data mining and had 
great success. Joe left IBM and joined a semiconductor startup in 
early 2001, while I left IBM earlier in late 1999. In his last computer 
chess work at IBM, Joe modified Deep Blue to provide a simple 



268 Chapter 13 

computation engine for the Deep Blue web server. Murray and Joe, 
as well as myself, have given presentations and written papers on 
Deep Blue. However, as far as I know, neither Murray nor Joe is 
working on computer chess any more. 

Changing Landscape 

In my opinion, it was Garry's own choice that he did not publicly 
play a computer match again. 

From what I know, IBM had high hopes for a good post-match rela- 
tionship with Garry. There were plans for possible commercial 
spots with him, whether Deep Blue won or not. His performance 
at the post-match press conference put those plans on hold. But as 
far as I can tell, to show good sportsmanship, IBM was still willing 
to grant him another match, at least initially. 

For a short while after the rematch, I worked on a more pow- 
erful version of the chess chip. The possibility of a new match and 
the desire to remove any doubt from the Kasparov camp that he 
lost fair and square, drove me on. Then, one day, it dawned upon 
me. "Get a life. You are free." In the twelve years of my life 1 had 
dedicated to the project I forgot there was a world out there. 

I knew that I wanted to do something else; the team had 
many public engagements after the rematch, and I did not have 
much time to think about what I really wanted to do for many 
months. Some time during this period, it became clear that there 
would not be any further IBM matches with Garry. There was no 
official decision, but the never-ending machinations finally took 
their toll. 

When I finally had some time to myself, I made three deci- 
sions. 

First, I decided to write this book. While Deep Blue is now a 
household name, few know about the real people behind Deep 
Blue. At least Murray and I are well known in computer chess cir- 
cles. Joe has the worst of it. Even computer chess people don't 
know much about him. As the person who instigated the project, 
I believe that it is my responsibility to set the record straight. 
Besides, I do have stories to tell. 

Second, I decided to get legal clearance from IBM to pursue 
the possibility of further developing the chess chip outside of IBM. 
Given that there would not be any IBM match with Garry, the only 
way a new match could take place would be for me to leave IBM, 



Life after Chess 269 

and build a new machine. If Garry really wanted a new match, I 
would do what I could to give him one. If he did not really want a 
new match, well, I would make an honest attempt to answer his 
challenges. 

Third, I would leave IBM once I regained control of the chess 
chip. I wanted to pursue some off-the-wall interests on my own. 
Seeing what Andreas was doing in his spare time reminded me 
how much fun it was trying to do things out of the ordinary. 
Maybe it is a character flaw of mine, but every once in a while I like 
to poke my head into research areas in which I have no known 
expertise. At Carnegie Mellon, I had to take remedials for two of 
the core qualifier courses, in Computer Systems (hardware 
design) 1 and Artificial Intelligence. Based on my course results, I 
had no business to be the system architect of a machine like Deep 
Blue. But that was part of the fun. I did not know where my off- 
the-wall ideas would lead, but it would be an adventure. Anyway, 
this was the real reason behind my leaving IBM. Licensing the 
chess chip from IBM also required that I left IBM after signing the 
contract. (It was the precondition before any licensing negotia- 
tion.) But I would have left IBM, whether the licensing deal went 
through or not. 

I left IBM in October 1999. After more than a year of contract 
negotiation, I finally had the legal clearance to do more or less 
whatever I wanted with the chess chips. I had saved up enough 
money 2 to build a PC-based chess machine that was more power- 
ful than Deep Blue if I absolutely had to. Of course, if possible, I 
preferred to keep the fund in reserve for my other interests. 

The Deep Blue chess chips were manufactured in 0.6 micron 
CMOS technology. It was possible to get 0.35-micron chips or even 
0.25 micron chips at reasonable cost by then. By going to more 
advanced technology, and by refining the design of the chess chip, 
it would have been possible to create a single chess chip 3 that 
would be almost comparable to the entire Deep Blue in chess play- 
ing strength. I would have to work like a mad man, but it could be 

1 Well, I was not that bad. I took the exam for the Computer Systems course 
directly without taking the course itself or studying for the exam. 

2 IBM would not have allowed me to license the chess chip if 1 did not have 
enough resources to build a new chess chip. 

3 The chip should run at 30-100 million positions/sec, depending on the tech- 
nology used and how many chess processors per chip. 



270 Chapter 13 

done. With a few such chips, a PC costing less than one tenth of 
Garry's one day appearance fee of $30,000, should be more than 
adequate to beat him in a match, assuming that the fixed engi- 
neering cost associated with fabricating the chip is ignored. 

Garry was still publicly making challenges for a new match 
and even went as far as suggesting that he would be willing to 
make the match a World Championship match. Personally, I 
didn't believe that machines should be allowed to play for the 
World Championship, but if Garry was willing, I was not going 
to argue with him. After all, his willingness would have made 
potential match sponsors far more interested. 

My first action was to gauge the interests of a few potential 
match sponsors. The responses I got were that there would be 
some interests especially if Garry was indeed willing to make it a 
World Championship match. The match would then be an his- 
toric first. However, despite Garry's public challenges, the 
potential sponsors expressed serious doubts that he would really 
agree to a match. It was clear to anybody with some knowledge of 
the situation that IBM would no longer agree to a new match. He 
must have known perfectly well that he did not have to worry 
about his challenges being answered, that is, until I left IBM. The 
other problem was that he could always say that the machine yet 
to be built was not Deep Blue, even though the new machine 
would have represented the only chance that he would ever have 
to play again against Deep Blue, or something close to it. 1 could 
not find any argument to refute the sponsors' reasoning. Deep 
down, I had some doubts of my own. The sponsors would not do 
anything until they were sure that he was willing to play. This 
went on for about two months. At this point, I promised myself 
that if he did not seem interested, I would begin to pursue my 
other interests. After all, the only reason I was going through the 
motion of trying to set up the match was to answer his challenges. 
If he would not take his own challenges seriously, why should I 
bother? 

Over the next month or so I made several contacts with 
Garry's manager, Owen Williams. Owen immediately ruled out 
any possibility of a World Championship match. This was fine by 
me. I was not comfortable with the idea of playing for the title. He 
was non-committal about the match; I could not even get him to 
say that, in principle, Garry was willing to play if a sponsor could 
be found. After a month of running around, I got final confirma- 



Life after Chess 271 

tion that Garry was not interested. I then sent out an open letter 
to a few widely read chess web sites and a copy to Owen. In the let- 
ter, I said my goodbyes, and thanks to the chess world, including 
Garry himself. I included information to say that Garry's challenge 
for a new match would be unanswered, since the challenge no 
longer existed. I did not say anything about why Garry was no 
longer interested in the match as Owen had not given me any hint 
as to what they were. 

A few days after my open letter, Owen published his response 
that listed several reasons why Garry was not interested in playing 
a match. Among the reasons was that I had no credibility, no 
money, and I was not serious. Anyway, it was a no-win situation 
for me to get involved any further. I posted a short reply, pointing 
out that I went to the trouble to secure the right to the chess chip 
and I had enough personal funds to build a new machine if I 
absolutely had to. I did not bother to mention that I began with 
no computer chess credentials in 1985 and we built our first 
machine (ChipTest) with no budget whatsoever, which did not 
stop Deep Blue from beating Garry in 1997. It was all a matter of 
will. Maybe there were genuine misunderstandings in the whole 
process and Garry was sincerely interested in a match. Anyway, the 
issue was no longer of concern to me. I had fulfilled my obligation 
to try to answer the challenge. 

There is nothing for Garry to gain by playing a new match, 
and it is very likely that he would be crushed. Computers can get 
better much faster than he can. In 1997, it took a multi-million- 
dollar supercomputer to beat him. In late 1999, a cheap PC in a 
college dorm with a few chess chips, fabricated in say 0.35 micron 
CMOS, should be able to do the job. During the 2001 Superbowl 
television broadcast, Pepsi ran a spot with Garry. In the commer- 
cial, he was seen beating a computer that looked like Hal (the 
insane computer in the novel/movie "2001") and then losing to a 
Pepsi vending machine. The reality in 2001 is that a palmtop com- 
puter like a Compaq iPaq, equipped with a 0.13 micron CMOS 
chess chip would have a decent chance to beat him. If you want to 
get fancy, you can even fit the entire machine into a cola can. It 
would be one very expensive can, but it can be done. 

With Garry presently titleless, who might be an interesting 
opponent for computers? My gut feeling is that Vladimir could be 
a tougher opponent for computers. Garry's playing strength 
resides to a great extent in his tactical play, which happens to be 



272 Chapter 13 




At a 1997 public exhibition in Taipei, Taiwan. 

the forte of chess computers. The way Vladimir neutralized 
Garry's tactical play in their match means that he probably can do 
the same thing to a computer, which will make him a very dan- 
gerous opponent indeed. 

A ten-game match between Vladimir and the program Deep 
Fritz was due to be played in October 2001 but was postponed. At 
the time of this writing, the match was rescheduled for October 
2002. In case you are wondering, no, Deep Fritz has nothing what- 
soever to do with Deep Blue. Ever since Deep Blue's win over Garry, 
several commercial programs have adopted the name Deep "what- 
ever" for their multiple-processor versions. Deep Fritz is 
supposedly able to search over a million positions/sec. The base 
program Fritz, has an evaluation function that is probably worse 
than Deep Thought's. This means that an alert human opponent 
will be able to exploit Fritz's positional weaknesses. Vladimir 
should win the match easily, unless he gets careless. 



What will be the next research area of interest in computer games? 
After chess, there is no clear choice. Among games with perfect 
information (games that do not depend on chance events like coin 
flips, die rolls or card shuffling), "Go" is generally recognized as the 
most difficult game. It is probably sufficiently difficult that it will 
not be "solved" within the next 20 years. Of the other games, two 
are of particular interest: Chinese chess and shogi (Japanese chess). 



Life after Chess 273 

Technically, Chinese chess is not very interesting, since its charac- 
teristics are very similar to chess. The number of Chinese chess 
players is possibly comparable to the number of chess players. If 
computers had been developed in China first, then this book 
might very well have been about a Chinese chess computer. To me, 
Chinese chess holds some allure — it was the first board game 1 
learned as a kid. But I doubt that I would spend time on it. It is just 
not sufficiently interesting technically. Shogi is an interesting case. 
It has a complexity between Go and chess, mainly from the rule 
that allows captured pieces to be dropped back onto the board. I 
have some ideas about how to tackle the computation complexity 
of shogi. Some day, I might build a shogi machine just to see how 
hard it is. It would not be a full-time effort like the Deep Blue proj- 
ect was — I want to do something practical for a change, and a 
full-time effort on another computer game is just not on the cards. 
I used to play Go passionately, but I don't believe I will ever work 
on it. The game is too hard for a computer at the moment. 

As far as I am concerned, computer chess is no longer a part 
of my life. Would I agree to build a new machine to play against 
Vladimir Kramnik or Vishwanathan Anand? I seriously doubt it, 
unless someone makes me an offer that I cannot refuse, as I am 
having too much fun with new projects. At this point in my life, it 
is far more important for me to do new and exciting things, rather 
than going back to an old conquest, exciting though it may be. Do 
I miss the thrill of competition? Sometimes. But it is time for me 
to get on with the rest of my life. The world is full of interesting 
things to do. 



i. 



Appendixes 



m 



APPENDIX A 



A Lad from Taiwan 



This is background material, mostly about my life in Taiwan. It 
should give you a sense of what kind of person I am at the core, 
and what kind of bias may exist in telling my stories. The 
Creatures in the Box describes my first contacts with chess. 
Learning Board Games details how I learned Chinese chess, Go, 
and chess. It also provides a short primer on computer chess his- 
tory up to the 1970s. An Early Riser is a short story about my 
early education, and an incident that I need to come clean 
about. Sidestepping Authorities provides some instances of my 
early tendency to ignore authority figures. This tendency has a 
bearing on what happened at Carnegie Mellon. The Case for 
Carnegie Mellon goes over the reasons why I ended up at 
Carnegie Mellon. 

The Creatures in the Box 

I was born in Keelung, Taiwan in 1959. Keelung is one of the major 
seaports of Taiwan, and at the time it had a population of around 
200,000. In Taiwan, Keelung is known as the "Rainy Port" because 
typically it rains for more than 200 days a year. It is a bustling city, 
but not a glamorous city by any stretch of the imagination. It is 
not a gloomy city either; I probably saw rainbows more often in 
Keelung than anywhere else in the world. 

When I was in kindergarten, our next-door neighbor bought 
the first television set in the neighborhood, it was a black and 
white set, as there was no color television broadcast in Taiwan yet. 
Every afternoon, the kids in the neighborhood would gather 
around to watch cartoons. I wondered how the creatures got into 
the box. There were a few small holes behind the box but they 

277 



278 Appendix A 

were too small for the creatures to get through. After a few weeks, 
the novelty wore off, and we got back to our normal games. 

A few years later, my dad bought a television set for the fam- 
ily. At the time, there was only one television station, and the local 
production of television shows was close to nonexistent. The sta- 
tion broadcast only on afternoons and evenings, but to fill the 
time slot, it ended up rebroadcasting American TV series, with 
Chinese captions. 

Chess was not big in the United States at the time. The boom 
resulting from Bobby Fischer winning the world title came a little 
bit later. But, occasionally, some of the foreign TV series carried 
episodes with chess themes. The first time I saw chess being played 
was in an episode of the "Family Affair" series. To me, it was a 
strange game played with beautifully sculptured pieces. 

In the early days, most of these imported TV shows were fam- 
ily sitcoms. As new stations began broadcasting, a wider variety of 
shows were introduced. Spy shows became quite popular. "Mission 
Impossible", "The Avengers", "Wild, Wild West", "it Takes a Thief, 
"The Man from U.N.C.L.E", "Get Smart", and so on. The political 
symbolism of the 1972 Fischer (USA) versus Spassky (USSR) World 
Chess Championship Match caught the attention of scriptwriters, 
and chess was incorporated into some of the story lines. 

One of the recurring TV themes went something like this. The 
bad guy is a Grandmaster chess player, usually of East European 
extraction, and very smart. Our hero has to play a chess match 
against the bad guy for some reason, perhaps to gain access to the 
bad guy, rescue another Grandmaster who is defecting, or stop 
something sinister at a chess competition. Obviously, it is prefer- 
able that our hero beat the bad guy, and thus upset him into 
making careless mistakes. There is one big problem. Even though 
our hero normally comes from the same country as Bobby Fischer, 
he is no chess Grandmaster. Of course, the USA is also known for 
its computer technology. So our hero is equipped with a well-hid- 
den remote hookup to a chess computer. The game against the bad 
guy usually starts off fine for our hero. For dramatic effect, some- 
thing typically goes wrong in the middle of the game. In the end, 
either the computer gets back on line and finishes off the bad guy, 
or our hero stumbles onto a move that checkmates the bad guy. 
The camera zooms into the face of the bad guy, who is in total dis- 
belief. Tho now chastened bad guy then storms out of the hall, and 
before he can set off the bomb or do something else nefarious, our 



A Lad from Taiwan 279 

hero catches up with him, and stops him cold, just in the nick of 
time. 

What is the moral of this story line? Well, computers are great 
tools for playing the game of chess against bad guys, especially if 
you are a handsome and debonair spy 1 . 

Learning Board Games 

Chess is not popular in Taiwan. Both Chinese chess and Wei-chi, 
which is better known as Go in the West from its Japanese name, 
are far more popular. Of the three, Chinese chess is the most pop- 
ular. It was also the first board game I learned as a kid. 

Now, if this book was about the life of some famous chess 
player, 1 would be telling you about how I learned by watching my 
dad and older brother play, and one day surprised them by point- 
ing out their mistakes in the game that they just finished. That 
would be a blatant lie. It was nothing like that. I learned the rules 
for Chinese chess on the street. 

There were very few cars in Keelung when I was little, and the 
kids in the neighborhood played all sorts of games on the streets. 
You might wonder how we could have played outside when it 
rained over half of the year. Well, the commercial buildings in 
Taiwan, especially in Keelung, were usually built two or three 
storeys high. The storefront was usually about ten feet from the 
street. The upper floor(s), supported by pillars next to the street, 
formed an overhang in front of the store. The overhang provided 
a nice shelter for the customers and the passers-by. You could walk 
from one end of a block of buildings to the other end without 
being rained on. The space under the overhang became our play- 
ground when it rained. 

I first learned what the pieces for Chinese chess were by 
watching older kids play the simpler variants of the game. Every 
once in a while, some of them decided to play "real" Chinese 
chess. Whenever they did, there was a sense of awe among us 
younger kids. The pieces move differently in the real game, and it 
took some time before I had an inkling of how the real game was 
played. My classmates in primary school were my earliest oppo- 
nents and my dad played some games with me when I got a little 
bit older. I became somewhat competitive with my dad after about 

1 Actually, they work well even if you are only an ordinary looking computer sci- 
entist and the opponent is a perfectly fine gentleman. 



280 Appendix A 

two weeks or so. One of my uncles visited and played some games 
with us. My uncle was far better than either of us, and afterwards 
graciously gave a pointer or two on how to start the game, and 
general playing principles. That was about the extent of my "for- 
mal" Chinese chess education. 

The playing strength of most of the Chinese chess players in 
Taiwanese schools apparently was quite weak. I was above average 
in playing strength when I was in junior high. One day after class, 
a friend introduced me to another Chinese chess player. I played a 
few games, and was completely outclassed. I was so discouraged 
that I played very few games thereafter. 

I did learn to play Western chess, or rather a bastardized form 
of chess, before I learned to play Go. My first game of "chess" was 
played against my elder brother, Feng-Lung. It was played on a 
homemade chess set that he put together. My brother won, of 
course, partly because of the fact he was telling me the rules as the 
game progressed. I learned the rule of pawn promotion 2 near the 
end of the game, when my brother promoted his pawn. This game 
was probably played when I was still in primary school. In junior 
high, some of my classmates would bring their magnetic chess sets 
to school once in a while, and we did play "chess" a little bit. None 
of us knew the rules well. In particular, our castling moves were all 
wrong. In Chinese, the castling move is translated as "swapping 
the castle", and we literally executed the move as a swap of the 
king with the rook. 

I started to get interested in Go when I was in junior high. As 
part of the school's after-school recreational programs, a local ama- 
teur 3-dan 3 player came in from time to time to teach us. We did 
not get much beyond the etiquette, rules and some basic patterns. 

2 A pawn reaching the 8th rank is promoted to a more powerful piece, usually a 
queen, the most powerful piece in the game. 

3 The ranking system in Go is a three-tiered system. Within each tier, players are 
ranked either by the number of kyu's or the number of dan's they have. At the 
lowest tier, the players are ranked by the number of kyu's, with stronger players 
having fewer kyu's. The one-kyu players are the strongest kyu-Ievel players. 
Above the kyu-Ievel, we have the amateur dan-level players, and the professional 
dan-level players. At dan-level, a player with more dan's is the better player. One- 
dan players, also known as shodans, are the weakest dan-level players. For both 
amateur and professional dan levels, the highest ranked players are 9-dans. The 
amateur 9-dans are usually no stronger than professional shodans. An amateur 
3-dan player roughly corresponds to a master level player in chess. 



A Lad from Taiwan 281 

When I got into Cheng-Kuo Senior High in Taipei, the capital, one 
of my elder sisters, Jing-Feng, bought me a Go set as a gift 4 . 
Professional 9-dan Lin Hi-Fong (Rin Kaihou in Japanese), the Go 
pro who hailed from Taiwan, was winning major Go titles in 
Japan, and Taiwanese newspapers were covering his title matches 
extensively. It was at this time that I discovered Go books. Keelung 
was not a backward city, but it was not as cultured as Taipei. There 
were few bookstores in Keelung, and none near our home. In 
Taipei, within walking distance from Cheng-Kuo Senior High, 
there was an entire street of bookstores. The Go books changed 
how I approached technical problems later on in life. Whenever I 
encounter a new problem, my first reaction is usually to dig up 
everything in the literature f can find about the problem. In the 
old days, this meant going to the library or bookstores. Of course, 
today I would have to add Internet searches to the mix. 
Sometimes, it is better to think about the problem first before look- 
ing it up in the literature, but old habits die hard. 

Studying Go was quite important in my early education. In 
the Far East, Go is more than just a game. It is an integral part of 
the culture. In Japan, businessmen sometimes would use Go 
proverbs to describe their business strategies. My Go study greatly 
influenced how I approached life. One of the most important les- 
sons I learned from Go was to pick my spot. In Go, it is frequently 
unwise to react to every single provocation from your opponent; 
sometimes, it is far more profitable to judiciously ignore the provo- 
cation and move to a new area. In other words, be yourself, and 
don't always follow the crowd. This attitude sometimes backfires 
very badly, but it serves me well overall, in Go games and in real 
life. 

Learning Go probably helped me understand some of the 
deeper concepts in chess. Sometimes when I asked good chess 
players about a specific formal concept that I learned from the Go 
literature, the answer frequently was "yes, that is an interesting 
idea, and I do use the concept intuitively". The concept, however, 
is usually not named in the chess literature. Go is a complicated 
game, but it has very simple rules. Perhaps the simplicity of Go 
rules makes it easier to describe the abstract concepts explicitly. 

I learned the proper rules for chess during my freshman year 

4 Cheng-Kuo is said to be the best public high school for boys in Taiwan and quite 
difficult to get in. Hence the gift for achieving entrance. 



282 Appendix A 

in National Taiwan University. One day, I decided to look through 
the catalogue of the school library for books on chess and was sur- 
prised to find a computer chess book listed. I was in the Electrical 
Engineering Department, so computers were of some interest to 
me. Finding a computer chess book listed was quite exciting, espe- 
cially with the catalogue listing the author as none other than the 
World Chess Champion Mikhail Botvinnik. I was very annoyed 
when I could not locate his book. 

With hindsight, not finding the book might have been a 
blessing in disguise. The approach he advocated appeared to go 
nowhere. For twenty years or so, he wrote many articles on how 
well his program performed in test positions. But, even to this day, 
no outside researcher has seen his program playing a single move, 
let alone competing against other programs. Some researchers even 
accused him of falsifying his results in the test positions. He had 
some good ideas, at least according to the chess players to whom I 
have spoken, but the ideas were difficult to work with even on the 
more powerful Western computers. Behind the iron curtain, they 
were likely impossible to implement. He died in the early 1990s, 
right around the time when the personal computer became wide- 
ly available in the former USSR. 

Anyway, with my interest piqued, I went through all the jour- 
nal articles that I could find on computer chess. 1 learned the rules 
of chess, not from a book, but from one of the computer chess arti- 
cles that happened to give a detailed description of the rules. 

1 continued to check the library catalogue from time to time. 
During my sophomore year, a new computer chess book arrived in 
the school library. The book, a collection of papers from a com- 
puter chess workshop, was Chess Skill in Man and Machine, which is 
now considered a classic by computer chess researchers. The most 
important chapter in the book is on the chess program Chess 4.5 
by David Slate and Lawrence Atkin. (Most of the modern chess 
programs are in one way or another clones of the Chess 4.5 pro- 
gram.) At the time I read the book, I had no idea how influential 
it was to become. I was mostly surprised by its pessimistic tone, 
especially given the usually rosy reports in the press but the pes- 
simism turned out to be far more realistic. 

The computer chess literature generally regards Claude 
Shannon's 1949 lecture as the beginning of modern computer 
chess history. Some of the early researchers were quite optimistic. 
In 1957, Herbert Simon, a Nobel laureate in Economics, and 



A Lad from Taiwan 283 

renowned computer scientist, went as far as stating "within 10 
years a digital computer will be the world's chess champion, unless 
the rules bar it from competition". Ten years later, the best pro- 
gram was authored by Richard Greenblatt, an MIT student. His 
program played at the low end of Class C level which, on the chess 
rating scale, would be 1400 to 1600 rating points. World 
Champions are typically at around 2800 to 2900 rating points on 
the same scale. A 1400 player is better than a beginner, but would 
lose to more than half of the weekend tournament players, let 
alone the World Champion. A player one class stronger (200 rating 
points higher) than their opponent should win about three games 
out of every four. The best computer chess program in 1967 was 
playing at 1400 points, or seven classes, below the World 
Champion. Herbert Simon's prediction was way off the mark. 

In 1975, when most of the material in Chess Skill in Man and 
Machine was written, Chess 4.5 was the best program. Its program- 
mers, however, were consistently maintaining that the program 
was still at Class C level, or no more than 200 points higher than 
Richard Greenblatt's. In other words, the best chess programs 
gained less than 200 points after almost 10 years of development. 
At this rate of progress, it would have taken 60 more years, or 
around the year 2035, before a computer could challenge the 
World Champion. No wonder the general tone of the book was 
pessimistic. Some optimism came back a few years later when 
Chess 4.9, running on a supercomputer, was shown to play at close 
to 2000 rating points. But I did not know about the renewed opti- 
mism until I was studying for my PhD degree at Carnegie Mellon 
University many years later. 

Chess books were difficult to find in Taiwan. The ones avail- 
able were in English and I had to go to specialty bookstores to find 
them. Those that I found were either game collections or intro- 
ductory books. The introductory books were too elementary — they 
did not go much beyond the game rules, and the game collections 
did not explain things at my level. My first contacts with chess 
positional concepts were from the computer chess literature. It was 
only when I was in the United States that I read my first book on 
chess positional concepts. I was a very weak chess player when I 
moved to the United States. 

During my college years, I continued to play Go from time to 
time. I also played a few games of chess, mostly because of my 
interest in computer chess. I hardly ever played Chinese chess, 



284 Appendix A 

although I do remember winning one casual game when I tried to 
"think like a computer", by imagining how a program would "rea- 
son" about the position. My thought process in that Chinese chess 
game, however, had no bearing whatsoever on how Deep Blue 
actually played chess decades later. 

After college, I spent two years in mandatory military service, 
and then went to Carnegie Mellon University in the United States. 
I quit playing Go when I was in the US. I was about shodan (one 
dan) on the American scale when I quit. I started my own com- 
puter chess project at Carnegie Mellon about a year after quitting 
Go, but the two decisions were totally unrelated. I was never real- 
ly a chess player, and played few games even after 1 began working 
full time on computer chess. I became better at chess because of 
the computer chess work, but I have never had a formal chess rat- 
ing. I am pretty certain that I was, at best, no better than an 1800 
on the chess rating scale. What I really learned from the Deep Blue 
project is a stronger appreciation of the beauty of the art practiced 
by the Grandmasters. 

An Early Riser 

The following story has nothing to do with chess or computer 
chess, but it does tell you something about me. The reason why I 
include it here is to come clean. My parents, my teacher and my 
classmates in junior high deserve to know the truth. 

The education system in Taiwan can be traced back to the 
Tang Dynasty in China. During this period, an Empress took over 
the reign of China after the death of her Emperor husband, becom- 
ing the first and only Empress to directly rule over the country. 
While traditionalists might not look upon her reign favorably, one 
of her edicts had profound effects on the Chinese psyche, and the 
family tradition of the entire Far East. She established a system for 
selecting government officials based on public exams. The public 
exams became the main social migration paths as there was no 
restriction on who could take part. Potentially, the son of the 
poorest peasant could become the Prime Minister. Education of 
children became the most important thing in family life for many 
parents in the Far East as other rulers in the region copied the sys- 
tem. In time, the education system itself began to revolve around 
examinations of various sorts. This thousand-year-old education 
system was designed to produce obedient civil officials for the 
imperial government. Creativity was, of course, not necessarily a 



A Lad from Taiwan 285 

desirable trait for obedient civil officials. 

At the time when I was enrolled in primary school, a 
Taiwanese youngster needed to go through three entrance exams 
before reaching college. The first entrance exam was for junior 
high (after six years of primary school), the second was for senior 
high, and the last was the college entrance exam. The first two 
were regional, but the college entrance exam was nationwide. 
During my third grade, the Taiwanese government decided that all 
children should have at least nine years of education, and the jun- 
ior high entrance exam was abolished. So I did have a little bit of 
childhood without the immediate pressure of the entrance exams. 

The entrance exams placed a premium on memorization, and 
the whole education system in Taiwan was seriously skewed as a 
result. Many students in Taiwan would spend time in private tutor- 
ing classes after regular school, studying for the entrance exams at 
considerable expense to their parents. To tell the truth, I have no 
idea what or how they studied in those private classes, as I never 
attended one myself. I managed to avoid taking private classes 
because I developed a peculiar method of "studying" while I was in 
primary school. My way of "studying" was only good for memo- 
rization, and not for truly understanding the subject. A few 
subjects, such as math and physics, still required regular study, but 
1 pretty much got by with the other subjects that only required 
memorization. 

My brother was in junior high while I was still in primary 
school. Occasionally, he would get up very early on the day of 
school exam because he had not finished studying from the day 
before. But when I thought about it, it made perfect sense to do the 
studying on the very day of the exam, assuming that it could be 
done. There were several good reasons. 

First, my memory would still be fresh, without the decay from 
a night of sleep. Second, the short time available for studying 
would force me to focus, and thus less studying time was actually 
needed. Third, given Taiwan's subtropical weather, it was far more 
pleasant to study in the early morning. Of course, my true moti- 
vation was that if the scheme worked, I would not need to spend 
a lot of time studying. I am a lazy person, and my laziness started 
very early in life. 

By the time I was in junior high, early morning cramming 
had become a ritual for me. On the day of a school exam, I would 
get up at, say 1.00 am, and study until daybreak. After taking the 



286 Appendix A 

exam, I crashed and forgot all the silly tidbits that I force-fed into 
my brain in the early morning. I was getting fairly good grades, 
constantly at or near the top of the class. My teacher started to take 
notice, as I was the only student at the top of the class not taking 
private classes. 

One morning, my teacher gave a speech to the entire school. 
That morning I was on duty to clean up the classroom, and thank- 
fully was not present for the speech. I was shocked to find out later 
that I was the subject of part of the speech. Apparently, my teacher 
got wind of the fact that I was studying in the wee hours of the 
morning, and gave a glowing account of how the other students 
should follow my example. He was assuming that I did it every 
day, and was using me as an example on how the other students 
should study as hard as I supposedly did. 

My teacher never did realize the real reasons for my early 
morning studying. I was even more embarrassed when my mother 
proudly repeated what my teacher said after hearing it from my 
classmates' mothers. I never did tell my mother or my teacher 
what really transpired. Well, now they know. 

Sidestepping Authorities 

Teenagers are known to be rebellious. I was reasonably well 
behaved, but I did have a rebellious streak. I respected the author- 
ities when they proved to my satisfaction that they deserved 
respect, but respecting the authorities did not mean that I would 
not try to sidestep them. This attitude towards authoritative figures 
had some bearing on some of the later events at Carnegie Mellon, 
during the early years of the chess project 

My brother was studying Chemical Engineering in college 
when I enrolled in senior high, and I picked up some of his college 
textbooks that he was no longer using. My first math teacher at the 
senior high was a fairly irresponsible person, who spent more than 
half of the time in class complaining about this and that instead of 
doing real teaching. After listening to him a few times, I decided 
that my time in class would be better spent doing my own reading, 
so I started going through my brother's math textbooks. The math 
teacher was so busy complaining about his life that he never real- 
ized that I was reading my own books, or maybe he just did not 
care. My second math teacher was a good teacher, but since I had 
already gone through some of the material, I continued my own 
extracurricular math study in the class from time to time. One day 



A Lad from Taiwan 287 

the teacher noticed that I was not paying attention to him and 
came down on me. When he found out what I was reading, he 
backed off but gave me a warning not to do it again. Not surpris- 
ingly, he began to pick on me, frequently asking me to go up to the 
blackboard and demonstrate that I knew the subject. This went on 
for a while until he came up with a problem that he thought I 
could not possibly solve. 

It turned out to be solving a second order finite difference 
equation 5 resulting from a random walk problem. Of course, finite 
difference equations are not really part of the high school curricu- 
lum. As a matter of fact, they were not in my college curriculum 
either. But I got lucky. They were in my brother's college curricu- 
lum, and I happened to have read about them a few months earlier 
in his books. I went up to the blackboard and solved the equation 
in about 30 seconds, using notations none of the other students 
knew anything about. It was priceless to see the teacher's face. 1 
knew that he could not possibly have expected me to solve the 
equation using the method that 1 used. I was curious about how he 
was going to explain things to the class, and what his alternative 
solution was. 

"Well, that was called a finite difference equation, but we are 
not going to solve it that way." So he did have an alternative way 
to solve the equation. This was going to be interesting. For the next 
ten minutes or so, he filled up half of the blackboard with compli- 
cated algebraic equations, and eventually obtained the same 
solution as I had earlier. I checked the supplementary books after- 
wards, and found out that was indeed the solution recommended 
in them. 

The teacher eased up on me after the incident. I made quite 
an impression on some of my high school classmates, and they 
gave me the nickname "Crazy Bird 6 ." Later on, when I came to the 
United States, people who had problems pronouncing my name 
knew me as CB. 

After 1 enrolled in the National Taiwan University, I had a few 
small incidents with some professors, but nothing serious. One 

5 If you don't know what a finite difference equation is, don't worry about it. I 
found the knowledge useful, but you certainly can read the rest of the book 
without knowing what it is. 

6 In Chinese, the word "crazy" has the same sound as "feng", the Chinese char- 
acter for "peak" and the first character of my given name. 



288 Appendix A 

incident took place in the physics class and was somewhat similar 
to what happened in my high school math class, but not as dra- 
matic. Afterwards, our physics professor had the strange notion 
that I would make a good physicist. I got good grades, so I had no 
reason to complain. I did something more daring in my Chinese 
class. We were supposed to write a Chinese essay on what we read 
in Chuan-Tzu, a Chinese literary classic. Chuan-Tzu happened to 
rhyme with the Chinese translation for "atomic". The title for the 
essay was something like "On Reading Chuan-Tzu". I changed the 
title to "Experiments on Chuan-Tzu Collisions" and wrote the 
essay as a satire in the form of a report on atomic collisions exper- 
iments. In my essay, I made fun of our Chinese professor 
extensively. He was not too happy with my portrayal of him, but 
was sufficiently amused to give me my highest ever grade for an 
essay. As I suspected, he did have a strong sense of humor. 

In Taiwan, boys were required to serve two years in the mili- 
tary after college. In my opinion, the net effect was that college 
boys became dumber; I certainly felt dumber. My reaction time was 
slower, and I definitely lost an edge along the way. The military 
environment was designed to produce obedient soldiers, not free 
thinking souls. The lack of intellectual stimulation led to mental 
atrophy. The fact that many of us were placed in high-pressure 
environments also did not help. I was luckier than most, in that 
my service was mainly teaching new soldiers and junior officers, 
the subjects of basic electricity and English, at the Missile 
Command Training Center. The hours were long, and it was frus- 
trating that the majority of the soldier students were from the 
bottom of the barrel. Most of the students either could not or 
would not learn the English alphabet. In the end, I just concen- 
trated on the few that had some chance of making it. Anyhow, I 
was not happy, and I wanted to change things. 

The Missile Command had a few simulators for training offi- 
cers in charge of firing the missiles. The simulators were expensive, 
and they were all broken. Before I became an instructor, I went 
through maintenance training for the command module, and was 
somewhat familiar with its internal structure. The simulator was 
based on a modified command module, and was effectively a glo- 
rified video game that used the radar screen as the display. The 
Center had the schematics for both the simulator and the com- 
mand module. They were based on ancient technologies: 
relay-based logic, analog potentiometers for waveform generation, 



A Lad from Taiwan 289 

and so on. Typical military stuff, ten to twenty years out of date. 
The video stream for the command module was easily accessible, 
and modifying it was trivial. To be able to turn a command mod- 
ule into a simulator was clearly desirable — there were many 
working command modules. I made a proposal to the Commander 
of the Training Center, and got the go-ahead to do a demonstra- 
tion. I went to the electronic-component stores in Taipei and 
acquired a few parts. It took about a day to put things together, and 
I was able to inject my own video signal onto the radar screen. The 
Commander was thrilled and I was asked to write a formal pro- 
posal. It was turned down in the end, as I had absolutely no 
intention of staying on beyond my regular service. 

Military service was a traumatic experience for me. Physically, 
I was fine. Psychologically, the long period of mental suppression 
took its toll. It was probably close to two years after military serv- 
ice before my mental acumen fully recovered. I managed to do a 
few things out of the ordinary while in the military, but being sub- 
ject to two years of unconditional authority was simply too much 
for me. Looking back, I wonder whether my early escapades in 
graduate school were a result of long deprivation of free will dur- 
ing my military service. 

My last year of military service was 1982. Early in that year I 
received notification from Carnegie Mellon University that I was 
accepted into their Computer Science Department. So, that 
August, a couple of weeks after leaving military service, I got on the 
plane and left for the United States. 

The Case for Carnegie Mellon 

How did I end up at Carnegie Mellon? Well, it was kind of convo- 
luted. 

The Taiwanese information technology (IT) industry was vir- 
tually nonexistent when I was an undergraduate in the Electrical 
Engineering (EE) Department of National Taiwan University. The 
microprocessor revolution had just begun. Seeing that the micro- 
processor might provide a way for the Taiwanese IT industry to 
leapfrog, the National Science Council in Taiwan funded micro- 
processor research projects in universities. As a result, in my 
sophomore year I began to work for our department head on a 
microprocessor research project. 

I joined the project after a classmate of mine, Chia-Min 
Hwang, asked whether 1 was interested in partnering him to work 



290 Appendix A 

on the project. For this project, opposite to what happened in the 
chess project, I was the software guy and Hwang was the hardware 
guy. Our project was to do vibration analysis of tooling machines. 
The project lasted until my graduation. In retrospect, my work was 
not much in terms of real research, but I developed some useful 
computer skills. This experience played a role in my choice of grad- 
uate school. 

The university-based microprocessor research projects were 
almost all software projects. Research in IT hardware was effective- 
ly out of the reach of Taiwanese universities. To handle hardware 
research, the government set up industrial research agencies, one 
of which was the Electronics Research and Services Organization 
(ERSO). In my senior year, a representative from ERSO gave a talk 
to students in the EE Department. He stressed that silicon chips 
would be critical in every aspect of our lives and, given the lack of 
natural resources in Taiwan, developing an IC industry would be 
vital for the future growth of the economy. The speaker from 
ERSO, Chintay Shih, is today the President of the Industrial 
Technology Research Institute, which includes ERSO itself. The talk 
must have been a good one, because afterwards I decided that I 
wanted to study how to design chips 7 . 

At the time the best place to study chip design was the United 
States, so I sent several applications to US universities. Most of my 
applications were sent to EE Departments. However, 1 did send one 
to the Computer Science (CS) Department at Carnegie Mellon 
University in Pittsburgh. There were two reasons that I sent my 
application there. 

The first reason was that my microprocessor experience made 
the idea of a CS department palatable, as long as I got to work on 
chip design. The fact that the CS Department at Carnegie Mellon 
was ranked among the top three did not hurt either. 

The second reason had something to do with my faculty advi- 
sor at Carnegie Mellon, Professor H T Kung, but not in the way you 
might think. I did not choose Carnegie Mellon because of Professor 
Kung's presence there; it was a bit more complicated. In the late 
1970s and early 1980s, Professor Kung did what he could to help 
the budding IC industry in Taiwan to get on its feet. One of the 

7 Well, I am exaggerating a little bit. 1 had an existing interest in learning how 
integrated circuits work. A personal project of mine forced me to learn about the 
subject. 



A Lad from Taiwan 291 

things that he did was to introduce the book Introduction to VLSI 
Systems by Carver Mead and Lynn Conway, to Taiwanese 
researchers. This book was the bible for VLSI classes in the early 
1980s. In the book, there was a chapter by Charles Leiserson (then 
a student of Professor Kung's) and Professor Kung himself. Charles, 
by the way, is now a professor at MIT. Anyway, from this chapter it 
was obvious that there was an active and well-respected VLSI group 
at the CS Department of Carnegie Mellon. This was the second rea- 
son for my sending in the application. 

I received two admissions from schools that I was really inter- 
ested in. One was for the EE Department at Stanford University, in 
the heart of the Silicon Valley. The other one was for the CS 
Department at Carnegie Mellon University. With my electrical 
engineering background, Stanford would have been my first 
choice, but the research assistantship offered by Carnegie Mellon 
tipped the scale, and I went to Pittsburgh 



APPENDIX B 



Selected Game Scores 



Chiptest-Recom, 1986 ACM Championship, Dallas, Round 5 

1. f4 d5 2. e3 c5 3. Bb5+ Bd7 4. Be2 g6 5. Nf3 Bg7 6. c4 dxc4 7. Bxc4 
b5 8. Be2 Bc6 9. O-O c4 10. a4 a6 11. axb5 axb5 12. Rxa8 Bxa8 13. 
Na3 Bxf3 14. Rxf3 Qd5 15. b3 cxb3 16. Bxb5+ Nd7 1 7. Bc4 Qb7 18. 
Bxb3 Ngf6 19. Rfl O-O 20. Nc4 Ra8 21. Ne5 e6 22. Nd3 Ne4 23. 
Nf2 Nxf2 24. Kxf2 Nc5 25. Bc4 Na4 26. Qc2 Nb6 27. Bb2 Nxc4 28. 
Bxg7 Nxd2 29. Qxd2 Kxg7 30. Kgl Ra4 31. Qc3+ f6 32. Rcl Qe4 33. 
Qb3 Ra7 34. Rc8 e5 35. Rc4 Qa8 36. Kf2 exf4 37. exf4 Ra2+ 38. Rc2 
Qa7+ 39. Kfl Ral+ 40. Ke2 Qe7+ 41. Kd2 Qd6+ 42. Ke2 Qxf4 43. 
h.3 Qe4+ 44. Kf2 Qh4+ 45. g3 Qxh3 46. Rc7+ Kh6 47. Qe3+ g5 48. 
Rxh7+ Kxh7 49. Qe7+ Kg6 1/2-1/2. 

Bent Lar sen-Deep Thought, 1988 Software Toolworks Chess 
Championship, Long Beach, Round 3 

1. c4 e5 2. g3 Nf6 3. Bg2 c6 4. Nf3 e4 5. Nd4 d5 6. cxd5 Qxd5 7. 
Nc2 Qh5 8. h4 Bf5 9. Ne3 Bc5 10. Qb3 b6 11. Qa4 O-O 12. Nc3 b5 
13. Qc2 Bxe3 14. dxe3 Re8 15. a4 b4 16. Nbl Nbd7 17. Nd2 Re6? 
(Ne5!, with the idea of Nf3 giving up the c pawn for a strong 
attack, possibly with Bd3 later) 18. b3 Rd8 19. Bb2 Bg6 20. Nc4 Nd5 
21. O-O-O N7f6 22. Bh3 Bf5 23. Bxf5 Qxf5 24. f3 h5 25. Bd4 Rd7 
26. Kb2 Rc7 27. g4 hxg4 28. Rhgl c5 29. fxg4 Nxg4 30. Bxg7 Rg6 
31. Qd2 Rd7 32. Rxg4 Rxg4 33. Ne5 Nxe3 34. Qxd7 Nxdl 35. Qxdl 
Rg3 36. Qd6 Kxg7 37. Nd7 Re3 38. Qh2 Kh7 39. Nf8 Kh8 40. h5 
Qd5 41. Ng6 fxg6 42. hxg6 Kg7 43. Qh7 Kf6 0-1 



292 



Selected Game Scores 293 

Deep Blue-Carry Kasparov, 1996 ACM Chess Challenge, Philadelphia, 
Came 1 

l.e4 c5 2.c3 d5 3.exd5 QxdS 4.d4 Nf6 5.Nf3 Bg4 6.Be2 e6 7.h3 Bh5 
8.0-0 Nc6 9.Be3 cxd4 10.cxd4 Bb4 ll.a3 Ba5 12.Nc3 Qd6 13.Nb5 
Qe7 H.NeS Bxe2 15.Qxe2 0-0 16.Racl Rac8 17.Bg5 Bb6 18.Bxf6 
gxf6 19.Nc4 Rfd8 20.Nxb6 axb6 21.Rfdl f5 22.Qe3 Qf6 23.d5 Rxd5 
24.Rxd5 exdS 25.b3 Kh8 26.Qxb6 Rg8 27.Qc5 d4 28.Nd6 f4 
29.Nxb7 Ne5 30.Qd5 f3 31.g3 Nd3 32.Rc7 Re8 33.Nd6 Rel+ 34.Kh2 
Nxf2 35.Nxf7+ Kg7 36.Ng5+ Kh6 37.Rxh7+ 1-0 

Garry Kasparov-Deep Blue, 1996 ACM Chess Challenge, Philadelphia, 
Came 2 

l.Nf3 d5 2.d4 e6 3.g3 c5 4.Bg2 Nc6 5.0-0 Nf6 6.c4 dxc4 7.Ne5 Bd7 
8.Na3 cxd4 9.Naxc4 Bc5 10.Qb3 0-0 ll.Qxb7 NxeS 12.Nxe5 Rb8 
13.Qf3 Bd6 14.Nc6 Bxc6 15.Qxc6 e5 16.Rbl Rb6 17.Qa4 Qb8 
18.Bg5 Be7 19.b4 Bxb4 20.Bxf6 gxf6 21.Qd7 Qc8 22.Qxa7 Rb8 
23.Qa4 Bc3 24.Rxb8 Qxb8 25.Be4 Qc7 26.Qa6 Kg7 27.Qd3 Rb8 
28.Bxh7 Rb2 29.Be4 Rxa2 30.h4 Qc8 31.Qf3 Ral 32.Rxal Bxal 
33.Qh5 Qh8 34.Qg4+ Kf8 35.Qc8+ Kg7 36.Qg4+ Kf8 37.Bd5 Ke7 
38.Bc6 Kf8 39.Bd5 Ke7 40.Qf3 Bc3 41.Bc4 Qc8 42.Qd5 Qe6 43.Qb5 
Qd7 44.Qc5+ Qd6 45.Qa7+ Qd7 46.Qa8 Qc7 47.Qa3+ Qd6 48.Qa2 
f5 49.Bxf7 e4 50.Bh5 Qf6 51.Qa3+ Kd7 52.Qa7+ Kd8 53.Qb8+ Kd7 
54.Be8+ Ke7 55.Bb5 Bd2 56.Qc7+ Kf8 57.Bc4 Bc3 58.Kg2 Bel 
59.Kfl Bc3 60.f4 exf3 61.exf3 Bd2 62.f4 Ke8 63.Qc8+ Ke7 64.Qc5+ 
Kd8 65.Bd3 Be3 66.Qxf5 Qc6 67.Qf8+ Kc7 68.Qe7+ Kc8 69.Bf5+ 
Kb8 70.Qd8+ Kb7 71.Qd7+ Qxd7 72.Bxd7 Kc7 73.Bb5 Kd6 1-0 

Deep Blue-Carry Kasparov, 1996 ACM Chess Challenge, Philadelphia, 
Came 3 

l.e4 c5 2.c3 d5 3.exd5 QxdS 4.d4 Nf6 5.Nf3 Bg4 6.Be2 e6 7.0-0 Nc6 
8.Be3 cxd4 9.cxd4 Bb4 10.a3 Ba5 ll.Nc3 Qd6 12.Ne5 Bxe2 
13.Qxe2 Bxc3 14.bxc3 NxeS 15.Bf4 Nf3+ 16.Qxf3 Qd5 17.Qd3 Rc8 
18.Rfcl Qc4 19.Qxc4 Rxc4 20.Rcbl b6 21.Bb8 Ra4 22.Rb4 Ra5 
23.Rc4 0-0 24.Bd6 Ra8 25.Rc6 b5 26.Kfl Ra4 27.Rbl a6 28.Ke2 h5 
29.Kd3 Rd8 30.Be7 Rd7 31.Bxf6 gxf6 32.Rb3 Kg7 33.Ke3 e5 34.g3 
exd4+ 35.cxd4 Re7+ 36.Kf3 Rd7 37.Rd3 Raxd4 38.Rxd4 Rxd4 
39.Rxa6 1/2-1/2 



294 Appendix B 

Garry Kasparov-Deep Blue, 1996 ACM Chess Challenge, Philadelphia, 
Came 4 

l.Nf3 d5 2.d4 c6 3x4 e6 4.Nbd2 Nf6 5.e3 Nbd7 6.Bd3 Bd6 7.e4 
dxe4 8.Nxe4 Nxe4 9.Bxe4 0-0 10.0-0 h6 ll.Bc2 e5 12.Rel exd4 
13.Qxd4 Bc5 14.Qc3 aS 15.a3 Nf6 16.Be3 Bxe3 17.Rxe3 Bg4 18.Ne5 
Re8 19.Rael Be6 20.f4 Qc8 21. h3 b5 22.f5 Bxc4 23.Nxc4 bxc4 
24.Rxe8+ Nxe8 25.Re4 Nf6 26.Rxc4 Nd5 27.Qe5 Qd7 28.Rg4 f6 
29.Qd4 Kh7 30.Re4 Rd8 31.Khl Qc7 32.Qf2 Qb8 33.Ba4 c5 34.Bc6 
c4 35.Rxc4 Nb4 36.Bf3 Nd3 37.Qh4 Qxb2 38.Qg3 Qxa3 39.Rc7 Qf8 
40.Ra7 Ne5 41.Rxa5 Qf7 42.Rxe5 fxe5 43.Qxe5 Re8 44.Qf4 Qf6 
45.Bh5 Rf8 46.Bg6+ Kh8 47.Qc7 Qd4 48.Kh2 Ra8 49.Bh5 Qf6 
50.Bg6 Rg8 1/2-1/2 

Deep Blue-Garry Kasparov, 1996 ACM Chess Challenge, Philadelphia, 
Game 5 

l.e4 e5 2.Nf3 Nf6 3.Nc3 Nc6 4.d4 exd4 5.Nxd4 Bb4 6.Nxc6 bxc6 
7.Bd3 d5 8.exd5 cxd5 9.0-0 0-0 10.Bg5 c6 ll.Qf3 Be7 12.Rael Re8 
13.Ne2 h6 14.Bf4 Bd6 15.Nd4 Bg4 16.Qg3 Bxf4 17.Qxf4 Qb6 18.c4 
Bd7 19.cxd5 cxd5 20.Rxe8+ Rxe8 21.Qd2 Ne4 22.Bxe4 dxe4 23.b3 
Rd8 24.Qc3 f5 25.Rdl Be6 26.Qe3 Bf7 27.Qc3 f4 28.Rd2 Qf6 29.g3 
Rd5 30.a3 Kh7 31.Kg2 Qe5 32.f3 e3 33.Rd3 e2 34.gxf4 elQ 35ixe5 
Qxc3 36.Rxc3 Rxd4 37.b4 Bc4 38.Kf2 g5 39.Re3 Rd2+ 40.Kel Rd3 
41.Kf2 Kg6 42.Rxd3 Bxd3 43.Ke3 Bc2 44.Kd4 Kf5 45.Kd5 h5 0-1 

Garry Kasparov-Deep Blue, 1996 ACM Chess Challenge, Philadelphia, 
Game 6 

l.Nf3 d5 2.d4 c6 3x4 e6 4.Nbd2 Nf6 5.e3 c5 6.b3 Nc6 7.Bb2 cxd4 
8.exd4 Be7 9.Rcl 0-0 10.Bd3 Bd7 11.0-0 Nh5 12.Rel Nf4 13.Bbl 
Bd6 14.g3 Ng6 15.Ne5 Rc8 16.Nxd7 Qxd7 17.Nf3 Bb4 18.Re3 Rfd8 
19. h4 Nge7 20.a3 Ba5 21.b4 Bc7 22x5 Re8 23.Qd3 g6 24.Re2 Nf5 
25.Bc3 h5 26.b5 Nce7 27.Bd2 Kg7 28.a4 Ra8 29.a5 a6 30.b6 Bb8 
31.Bc2 Nc6 32.Ba4 Re7 33.Bc3 Ne5 34.dxe5 Qxa4 35.Nd4 Nxd4 
36.Qxd4 Qd7 37.Bd2 Re8 38.Bg5 Rc8 39.Bf6+ Kh7 40x6 bxc6 
41.Qc5 Kh6 42.Rb2 Qb7 43.Rb4 1-0 



Selected Game Scores 295 

Carry Kasparov-Deep Blue, IBM Kasparov vs. Deep Blue Rematch, 
Game 1 

l.Nf3 d5 2.g3 Bg4 3.b3 Nd7 4.Bb2 e6 5.Bg2 Ngf6 6.0-0 c6 7.d3 Bd6 
8.Nbd2 O-O 9.h3 Bh5 10.e3 h6 ll.Qel Qa5 12.a3 Bc7 13.Nh4 g5 
14.Nhf3 e5 15.e4 Rfe8 16.Nh2 Qb6 17.Qcl a5 18.Rel Bd6 19.Ndfl 
dxe4 20.dxe4 Bc5 21.Ne3 Rad8 22.Nhfl g4 23.hxg4 Nxg4 24.f3 
Nxe3 25.Nxe3 Be7 26.Khl Bg5 27.Re2 a4 28.b4 f5 29.exf5 e4 30.f4 
Bxe2 31.fxg5 Ne5 32.g6 Bf3 33.Bc3 Qb5 34.Qfl Qxfl+ 35.Rxfl h5 
36.Kgl Kf8 37.Bh3 b5 38.Kf2 Kg7 39.g4 Kh6 40.Rgl hxg4 41.Bxg4 
Bxg4 42.Nxg4+ Nxg4+ 43.Rxg4 Rd5 44.f6 Rdl 45.g7 1-0 

Deep Blue-Garry Kasparov, IBM Kasparov vs. Deep Blue Rematch, 
Game 2 

l.e4 e5 2.Nf3 Nc6 3.Bb5 a6 4.Ba4 Nf6 S.O-O Be7 6.Rel b5 7.Bb3 d6 
8.c3 O-O 9.h3 h6 10.d4 Re8 ll.Nbd2 Bf8 12.Nfl Bd7 13.Ng3 Na5 
14.Bc2 c5 15.b3 Nc6 16.d5 Ne7 17.Be3 Ng6 18.Qd2 Nh7 19.a4 Nh4 
20.Nxh4 Qxh4 21.Qe2 Qd8 22.b4 Qc7 23.Recl c4 24.Ra3 Rec8 
25.Rcal Qd8 26.f4 Nf6 27.fxe5 dxeS 28.Qfl Ne8 29.Qf2 Nd6 
30.Bb6 Qe8 31.R3a2 Be7 32.Bc5 Bf8 33.Nf5 BxfS 34.exf5 (6 
35.Bxd6 Bxd6 36.axb5 axb5 37.Be4 Rxa2 38.Qxa2 Qd7 39.Qa7 Rc7 
40.Qb6 Rb7 41.Ra8+ Kf7 42.Qa6 Qc7 43.Qc6 Qb6+ 44.Kfl Rb8 
45.Ra6 1-0 

Garry Kasparov-Deep Blue, IBM Kasparov vs. Deep Blue Rematch, 
Game 3 

l.d3 e5 2.Nf3 Nc6 3x4 Nf6 4.a3 d6 5.Nc3 Be7 6.g3 O-O 7.Bg2 Be6 
8.0-0 Qd7 9.Ng5 Bf5 10.e4 Bg4 11. f3 Bh5 12.Nh3 Nd4 13.Nf2 h6 
14.Be3 c5 15.b4 b6 16.Rbl Kh8 17.Rb2 a6 18.bxc5 bxc5 19.Bh3 
Qc7 20.Bg4 Bg6 21.f4 exf4 22.gxf4 Qa5 23.Bd2 Qxa3 24.Ra2 Qb3 
25.f5 Qxdl 26,Bxdl Bh7 27.Nh3 Rfb8 28.Nf4 Bd8 29.Nfd5 Nc6 
30.Bf4 Ne5 31.Ba4 NxdS 32.Nxd5 aS 33.Bb5 Ra7 34.Kg2 g5 
35.Bxe5+ dxeS 36.f6 Bg6 37. h4 gxh4 38.Kh3 Kg8 39.Kxh4 Kh7 
40.Kg4 Bc7 41.Nxc7 Rxc7 42.Rxa5 Rd8 43.Rf3 Kh8 44.Kh4 Kg8 
45.Ra3 Kh8 46.Ra6 Kh7 47.Ra3 Kh8 48.Ra6 1/2-1/2 



296 Appendix B 

Deep Blue-Garry Kasparov, IBM Kasparov vs. Deep Blue Rematch, 
Game 4 

l.e4 c6 2.d4 d6 3.Nf3 Nf6 4.Nc3 Bg4 5.h3 Bh5 6.Bd3 e6 7.Qe2 d5 
8.Bg5 Be7 9.e5 Nfd7 10.Bxe7 Qxe7 1 l.g4 Bg6 12.Bxg6 hxg6 13.h4 
Na6 14.0-0-0 O-O-O lS.Rdgl Nc7 16.Kbl f6 17.exf6 Qxf6 18.Rg3 
Rde8 19.Rel Rhf8 20.Ndl e5 21.dxe5 Qf4 22.a3 Ne6 23.Nc3 NdcS 
24.b4 Nd7 25.Qd3 Qf7 26.b5 Ndc5 27.Qe3 Qf4 28.bxc6 bxc6 
29.Rdl Kc7 30.Kal Qxe3 31.fxe3 Rf7 32.Rh3 Ref8 33.Nd4 Rf2 
34.Rbl Rg2 35.Nce2 Rxg4 36.Nxe6+ Nxe6 37.Nd4 Nxd4 38.exd4 
Rxd4 39.Rgl Rc4 40.Rxg6 Rxc2 41.Rxg7+ Kb6 42.Rb3+ Kc5 43.Rxa7 
Rfl+ 44.Rbl Rff2 45.Rb4 Rcl+ 46.Rbl Rcc2 47.Rb4 Rcl+ 48.Rbl 
Rxbl+ 49.Kxbl Re2 50.Re7 Rh2 51.Rh7 Kc4 52.Rc7 c5 53.e6 Rxh4 
54.e7 Re4 55.a4 Kb3 56.Kcl 1/2-1/2 

Garry Kasparov-Deep Blue, IBM Kasparov vs. Deep Blue Rematch, 
Game 5 

l.Nf3 d5 2.g3 Bg4 3.Bg2 Nd7 4.h3 Bxf3 5.Bxf3 c6 6.d3 e6 7.e4 Ne5 
8.Bg2 dxe4 9.Bxe4 Nf6 10.Bg2 Bb4+ ll.Nd2 h5 12.Qe2 Qc7 13x3 
Be7 14.d4 Ng6 15.h4 e5 16.Nf3 exd4 17.Nxd4 O-O-O 18.Bg5 Ng4 
19.0-0-0 Rhe8 20.Qc2 Kb8 21.Kbl Bxg5 22.hxg5 N6e5 23.Rhel c5 
24.Nf3 Rxdl+ 25.Rxdl Nc4 26.Qa4 Rd8 27.Rel Nb6 28.Qc2 Qd6 
29x4 Qg6 30.Qxg6 fxg6 31.b3 Nxf2 32.Re6 Kc7 33.Rxg6 Rd7 
34.Nh4 Nc8 35.Bd5 Nd6 36.Re6 Nb5 37.cxb5 RxdS 38.Rg6 Rd7 
39.Nf5 Ne4 40.Nxg7 Rdl+ 41.Kc2 Rd2+ 42.Kcl Rxa2 43.Nxh5 Nd2 
44.Nf4 Nxb3+ 45.Kbl Rd2 46.Re6 c4 47.Re3 Kb6 48.g6 Kxb5 49.g7 
Kb4 1/2-1/2 

Deep Blue-Garry Kasparov, IBM Kasparov vs. Deep Blue Rematch, 
Game 6 

l.e4 c6 2.d4 d5 3.Nc3 dxe4 4.Nxe4 Nd7 5.Ng5 Ngf6 6.Bd3 e6 
7.Nlf3 h6 8.Nxe6 Qe7 9.0-0 fxe6 10.Bg6+ Kd8 ll.Bf4 b5 12.a4 
Bb7 13.Rel Nd5 14.Bg3 Kc8 15.axb5 cxb5 16.Qd3 Bc6 17.Bf5 exf5 
18.Rxe7 Bxe7 19.c4 1-0 



APPENDIX C 



Further Reading 



On Deep Blue and Its Predecessors 

Deep Blue, January 2002, Artificial Intelligence, vol. 134, pp. 57-83. 

IBM's Deep Blue Chess Grandmaster Chips, Mar/Apr 1999, IEEE Micro 

Magazine, pp. 71-81. 

A Grandmaster Chess Machine, October 1990, Scientific American, 

pp. 44-50. 

Deep Thought, 1990, a chapter in the book Computers, Chess, and 

Cognition, Springer-Verlag, pp. 55-78. 

The web site http://www.chess.ibm.com is an archive of the 1997 

Deep Blue match web broadcast. Deep Blue game logs from the 

1997 rematch can be found on this web site as of this writing. 

On Computer Chess in General 

Chess Skills in Man and Machine, Springer-Verlag, is still a classic. 
The chapter on Chess 4.5, while dated, is still an essential primer 
on how to program a chess computer. 

Computers, Chess, and Cognition, Springer-Verlag, is a book on more 
recent developments up to the early 90s. It also contains some 
material on computer go. 

Journal of the International Computer Chess Association, is a source for 
latest developments in computer chess. For subscription informa- 
tion, see http://www.dcs. qmw.ac.uk./~icca /. The web site also 
contains interesting links. 

On Other Games 

One Jump Ahead: Challenging Human Supremacy in Checkers, 
Springer-Verlag, is a book on the creation of Chinook, a World 
Championship class checkers program. It also contains a different 

297 



298 Appendix C 

perspective on the computer chess world in the 1980s and the 
early 1990s. 

http://www,ff.iij4u, or.jp/~junl/csa/index e.html is the web site for 
the Computer Shogi Association. 

http://www.usgo.org/computer/ is a web site maintained by 
American Go Association on computer Go.